The riscfactors that influence the likelihood of belonging to Severe Mental Ill group in GGZ Rivierduinen

(1)

GGZ Rivierduinen

Department of Medical Informatics (UVA)

Scientific Research Project

Thesis 30 October 2020 Master Medical Informatics

Master Medical Informatics

T

he

riscfacTors

ThaT

influence

The

likelihood

of

belonging

To

s

evere

M

enTal

i

ll

group

in

ggZ

r

ivierduinen

An analytical study on patients with SMI

W

riTTenby

:

Z. i

brahiM

s

upervisor

l. b

ouMa

r.

van

h

arTen

T

uTor

d

r

. J.h. l

eopold

GGZ Rivierduinen

Department of Medical Informatics (UvA)

Scientific Research Project

Thesis 31 October 2020 Master Medical Informatics

Scientific research on:

The influence of patients with severe mental

illnesses on the capacity and financial burden of

GGZ Rivierduinen

An analytical study on patients with SMI in GGZ Rivierduinen

Written by Z. Ibrahim Supervisor L. Bouma R. van Harten Tutor Dr. J. H. Leopold

(2)

W

riTTenby

edical

i

nforMaTics

(uva)

a

MsTerdaM

uMc

M

eibergdreef

9 1105 aZ a

MsTerdaM

p

racTiceTeachingperiod

a

pril

2020 - o

cTober

2020

T

he

riscfacTors

ThaT

influence

The

likelihood

of

belonging

To

s

evere

M

enTal

i

ll

group

in

ggZ

r

ivierduinen

(3)

(4)

4

I would like to take this opportunity to thank and express my appreciation for R. van Harten for his assistance and for creating a place for me at GGZ Rivierduinen and connecting me to innumerable aspiring people who have been a great support: J. van Zoelen, M.Mutschelknauss, G. van Son, K.Smith, M. Hendriksen, B. Luteijn, J. van den Heuvel, K. Westendorp, E. Furth, R. de Winter en I. Siteur and many whom I have not have mentioned. For Dr. J.H. Leopold to have faith in me and my work and guiding me through the thesis.

My indebtedness and great acknowledgment also goes for people who I knew little but helped me a great deal: Lisa- Milou Bouma and Eva without you this work will not have been possible. I would also like to express my gratitude to Tim, for being there in my stressy moments. And Atif for keeping company and encouraging me throughout the thesis.

To my family: Mom, Dana, Mejda, Medya, Karzan, Kurdo and Benjamin, friends and others who in one way or the another shared their support, Thank you for your faith, warmth and countless love.

Above all, to the great Almighty the author of all knowledge and wisdom, thank you for this journey.

- Zana Ibrahim

Acknowledgements

(5)

Abstract

Background: Patients with Severe Mental Illnesses are thought to place pressure on the capacity and finances of mental healthcare centres, for several reasons. Previous research has demonstrated that SMI patients have an increased Length of Stay (LOS) in clinics, they are more likely to get admitted to Emergency Department and their numbers are unpredictable for healthcare centres. Patients with SMI are treated as part of specialized GGZ and have often have various types of disorders at the same time which makes the needed care complex. This study aims to investigate the risk factors that influence the care requirements of the Severe Mental Ill population, this study is performed in GGZ Rivierduinen. The main question that is attempted to be answered is:

What factors influence the likelihood of a higher DBC price (in other words, belonging to the Severe Mental ill group) at GGZ Rivierduinen?

Methods: Baseline data from the GGZ Rivierduinen were used, including a total of N=31582 unique DBC’s over a period of 36 months. Out of total DBCs, n=2039 were classified as having SMI and n=29543 as non-SMI. Only completed and invoiced DBC’s were included in the study. Characteristics studied included Length of stay, Total number of comorbidity, treatment hours, and admissions to Emergency department.

Results: The results demonstrated a significantly higher number of treatment hours in first 30 days in SMI group, p <0.001, OR =1.00 (95% CI: 1.000- 1.001), a higher admissions rate to Emergency department, , p <0.001, OR =2.13 (95% CI: 2.04-2.22), and lower chance of early discharge (13% vs. 24%; p <0.001) than in non-SMI group. The Length of Stay is also relatively higher in SMI group in comparison to non-SMI, p <0.001, OR =1.011 (95% CI: 1.010-1.013). A patient who belonged to Recovery ward or district ward, they were more likely to belong to the SMI group, than if they belonged to Policlinic.

Limitations: substance/drugs usage and mental retardation were not included as independent variables.

Conclusion: Large differences exist between SMI population and non-SMI on the following risk factors: the departments they resided in; their admissions to the Emergency Department; Total number of Comorbidities; number of treatment hours and their Length of Stay at the centre.

Future research could do a cross-healthcare centre examination, to see whether SMI populations are similar across healthcare centres in the Netherlands.

KEYWORDS

(6)

6

Samenvatting

Achtergrond: Patiënten met Severe Mental Illnesses (SMI) zouden om verschillende redenen druk uitoefenen op de capaciteit en financiën van geestelijke gezondheidszorg centra. Uit eerder onderzoek blijkt dat SMI-patiënten een langere verblijfsduur hebben in klinieken, er een grotere kans is ze worden opgenomen op de afdeling spoedeisende hulp. Hiernaast is het aantal SMI patienten onbekend en onvoorspelbaar. Patiënten met SMI worden behandeld als onderdeel van de gespecialiseerde zorg en hebben tegelijkertijd meerdere soorten aandoeningen, waardoor de benodigde zorg stijgt en complex wordt. Dit onderzoek heeft als doel de risicofactoren te onderzoeken die van invloed zijn op de zorgbehoefte van de Severe Mental Ill populatie in GGZ Rivierduinen. De belangrijkste vraag die hiermee bantwoordt moet worden, is: Welke factoren hebben invloed op de waarschijnkelijkheid van een hogere DBC-prijs (oftewel behorend tot de groep Severe Mental Ill) bij GGZ Rivierduinen?

Methode: Er is gebruik gemaakt van data van de GGZ Rivierduinen. In deze studie is een omvang van totaal N = 31582 unieke DBC’s over een periode van 36 maanden onderzocht. Van de totale DBC’s werd n= 2039 geclassificeerd als SMI en n= 29543 als niet-SMI. Alleen ingevulde en gefactureerde DBC’s zijn in dit onderzoek geincludeerd. De bestudeerde kenmerken waren onder andere de verblijfsduur, het totale aantal co-morbiditeit, de behandelingsuren en opnames op de spoedeisende hulp.

Resultaten: De resultaten lieten een significant hoger aantal behandelingsuren zien in de eerste 30 dagen voor de SMI-groep, p <0,001, OR = 1,00 (95% BI: 1,000-1,001), een hoger aantal opnames op de afdeling spoedeisende hulp, p <0,001, OR = 2,13 (95% BI: 2,04-2,22), en lagere kans op vroegtijdig ontslag (13% vs. 24%; p <0,001) dan in de niet-SMI-groep. De verblijfsduur in klinieken is ook relatief langer voor de SMI-groep in vergelijking met de niet-SMI groep, p <0,001, OR=1,011 (95% BI: 1,010-1,013). Een patiënt uit de Herstelafdeling of -Wijk, behoort eerder tot de SMI-groep dan een patiënt uit de polikliniek. Limitaties: drugsgebruik en zwakzinnigheid zijn niet geincludeerd als onafhankelijke variabelen.

Conclusie: Er bestaan grote verschillen tussen SMI-populatie en niet-SMI op de volgende risicofactoren: de afdelingen waar ze verblijven; hun opnames op de afdeling spoedeisende hulp; totaal aantal co-morbiditeiten; aantal behandeluren en hun verblijfsduur in klinieken. Toekomstige onderzoek zou een cross-healthcare in verschillende zorgcentra kunnen doen om te zien of SMI-populaties vergelijkbaar zijn in elke zorgcentra van Nederland.

(7)

List of Tables

T

able

1

19. T

able

2

20. T

able

5.4

25. T

able

3.1 T

able

3.2 T

able

3.3

able

4.1 T

able

4.2 T

able

4.3

44. T

able

4.4 T

able

4.5 T

able

7.1

49. T

able

7.2 T

able

7.3 T

able

7.4 T

able

7.5

50. T

able

7.6 T

able

7.7

51.

(8)

8

List of Figures

Figure 1|The SMI population 12.

Figure 2|The research approach is outlined 14.

Figure 3|The average price per unique costumer (APUC) 21.

Figure 5|Probability belonging to a group based on the treatment hours 22.

Figure 7|Probability belonging to a group based on the ED admissions 23.

Figure 8|Kaplan Meier curve for univariate SMI/ non-SMI 24.

Figure 10|Illustration of Neural network 26.

Figure 11|Neural network of the risk factors 26.

Figure 12|Garson’s relative importance algorithm 27.

Figure 4| Residual plots linear model 42.

Figure 6| Residual plots linear model ED admissions 44.

Figure 9| Schoenfeld test for proportionality for the variables in multivariate cox regression 46.

Figure 13| Kaplan Meier curve for univariate total comorbidity 47.

Figure 14| Kaplan Meier curve for univariate number of admissions to clinic 47.

Figure 15| Olden’s relative importance algorithm 49.

(9)

a

cknoWledgeMenTs

4 a

bsTracT

5 s

aMenvaTTing

6 l

isTof

T

ables

7 l

isTof

T

ables

8 i

nTroducTion

11

1.1. Definitions 11

1.2. Characteristics of SMI patients 12

1.3. The Impact of SMI Patients on Healthcare Centres. 13

1.4. Research Question & Relevance of the Present Study 13

1.5. Research approach 14

b

ackground

15

2.1. Mental healthcare in the Netherlands 15

2.2. GGZ Rivierduinen 16

M

eThods

17

3.1. Study Sample 17 3.2. Operationalization of variables 17 3.3. Data collection 18 3.4. Data Analysis 18

r

esulTs

19 4.1 p

reliMinary

r

esulTs

19

4.2 Descriptive Results 20

4.3 Number of treatment hours in the first thirty days 21

4.4 Number of admissions to the Emergency Department 22

4.5 Probability of early discharge 24

4.6 Influence of the risk factors on the total DBC price 26

d

iscussion

29

5.1. Implications 31 5.2. Limitations 31 5.3. Future research 32

c

onclusion

33 r

eferences

34 a

ppendixes

37

Appendix A 37 Appendix B 39

Appendix C: Descriptive Results 40

Appendix D: Number of treatment hours in first thirty days 42

Appendix E: Number of admissions to the Emergency Department 44

Appendix F: Probability of early discharge 46

Appendix G: Influence of the risk factors on the total DBC price 48

(10)

(11)

A little over one million adults in the Netherlands utilise mental healthcare services annually [1]. Additionally, 43% of the Dutch population experiences a mental disorder at a certain point of their life [1]. These mental disorders can have a range of indirect consequences, including heightened chances of poverty and poor housing conditions, as well as increased rates of substance abuse and comorbid disorders [2]. The combined direct and indirect complications resulting from mental disorders directly affect the financing of health care institutions and places pressure on their available capacity, in terms of number of staff, available beds and medical equipment. There is a growing body of literature that recognizes the importance of research on patients with Severe Mental Illnesses (SMI) in healthcare institutions, as they are thought to place additional pressure on the capacity and finances of mental healthcare centres, for several reasons. First, research has demonstrated that SMI patients have an increased Length of Stay (LOS) in clinics [3]. Second, their numbers are unpredictable for healthcare centres, and frequently have comorbid disorders that require simultaneous treatment or transfer between departments, creating a disjointed care plan. Third, SMI patients more frequently require treatment in Specialised Care Units due to their complexity and the intensiveness of their required care [4][5]. All in all, research has established that this particular group of patients expend around 4.7 billion euros a year, accounting for a significant chunk of the Netherlands’ overall mental healthcare expenditure [6].

To be able to provide the best possible care for patients with SMI, with better preparation for their treatments and thus lower Length of Stay, mental healthcare centres gather information and study patients with a diagnosis of SMI. This study aims to do exactly this, by investigating the factors that influence the care requirements of the SMI population in this case study in GGZ Rivierduinen.

Introduction

c

hapTer

1 1.1.1. Epidemiological definition of SMI

In various countries, the concept of a Severe Mental Illness (SMI) is defined differently. These concepts can however broadly be distinguished into two classes: SMI defined based on epidemiological diagnosis and degree of disability in society, or based on degree of disability and quantity of care required by this group. [7][8]

The epidemiological definition considers the diagnosis, comorbid disorders, degree of disability and presence of abnormal behavior in classifying a patient as having a Severe Mental Illness [7][9]. A considerable amount of academic literature has been dedicated to patients with SMI. These studies define patients with SMI’s as individuals with: recognizable psychotic, schizophrenic syndromes, patients presenting with potentially prodromal symptoms and who are “clinical high risk(CHR)” [10]. According to Lawrence and Kisely, SMI patients are entitled to higher use of healthcare, because of the high level of internal (medication) and external (outward treatment and help) support they require [11]. This definition includes patients which are in need of constant care and suffer from: Major Depression, Bipolar Disorder or Schizophrenia [4][7][11].

(12)

12 1.1.2. Definition of SMI’s according to Delespaul

SMI’s have gained an increased amount of attention from healthcare centres following the implementation of a new action plan developed in 2013 by the Council for Public Healthcare (CPH) and Social Council Development (SCD) [12]. This action plan indicated that the care needs of these patients should serve as a guideline for the care provision, supplementary to their diagnosis they have. This action plan outlined a new way in which to treat, guide and support people with SMI in the mental healthcare centres in the Netherlands [13].

Here, the definition given by Delespaul (2011), categorizes SMI patients as those who endure one of the earlier/ above mentioned disorders, but it can additionally include patients who have a diagnosis of a long-term psychiatric disorder and suffer from serious social limitations [8][13][14]. Patients with SMI are in long-term care. According to van Os and Delespaul, the SMI group is in constant risk of suffering a crisis and is in constant need of professional care [8][15]. SMI individuals have low rates of social participation rate. In other words, they have almost no social network or activity: around 80% of SMI report that they suffer from loneliness and a mere estimated 20% of this group have a paid job [16].

In short, the SMI group is defined here based on the amount of care they require, the amount of services they use, the complexity of their diagnoses and different types of care they require. The division between SMI patients and the rest of population is better elaborated and explained in the methods section.

This study has selected Delespaul’s definition as its criterion for defining patients with SMI’s. This is not to say that it is necessarily a superior definition, but it suits the purpose of this study, in the sense that its focus is on preparing a healthcare center, GGZ Rivierduinen, for its most complex patient group, rather than considering diagnostic classifications necessarily.

There are around 220,000 SMI patients in the Netherlands. Around 81% of these patients are treated in mental healthcare centres, while 78% of SMI patients are admitted to a specialised clinic at some point during their treatment [6]. The SMI population consists of various diagnostic groups according to Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). SMI patients consist of those with diagnoses of: Alcohol-Related Disorders (Substance-Alcohol-Related and Addictive Disorders); Anxiety disorders; Bipolar and Alcohol-Related Disorders; Major and Mild Neurocognitive Disorders; Neurodevelopmental Disorders; Depressive Disorders; Feeding and Eating Disorders; Substance-Related and Addictive Disorders; Personality Disorders; Somatic Symptom and Related Disorders; and Schizophrenia Spectrum and Other Psychotic Disorders [Appendix A] [17]. In Figure 1, it can be clearly seen that the largest group presented in the SMI population are patient with Schizophrenia, Depressive disorders, Personality disorders and Bipolar related disorders. [Figure 1].

1.2. Characteristics of SMI patients

Diagnosis of the SMI population of GGZ Rivierduinen. Note: The largest group within SMI consists of patients with Schizophrenia (26%), patients with Depression (18%), Personality disorders (13%) and Bipolar disorder (10%). This data is from a preliminary data analysis of GGZ Rivierduinen’s population.

(13)

Those suffering from SMI’s have significant difficulties in various areas of their functioning, including psychological, social and financial status [11]. Patients with SMI’s often have chronic problems and require constant care. The majority group of these patients also have cognitive impairments, lack familial support and motivation in socio-economic factors [10]. The risk factors that influence the occurrence of SMI’s and correlate with higher total DBC price have been explored in several studies, these included factors such as: education level; age; gender; psychiatric comorbidity and type of diagnosis [5][7][10].

Due to the high amount of investment required by care institutions as well as the considerable suffering that patients undergo, previous research has focussed on what factors can predict SMI care levels, defined as the total DBC price. First, gender has been shown in previous research to be a significant predictor of eventual SMI DBC costs. Women are marginally overrepresented in the SMI group, constituting around 53% of these patients. Second, it is expected that education level will be a significant predictor of DBC price, with lower education being expected to correlate with higher DBC price. Third, it is also believed that patients with various comorbid- diagnoses are correlated with DBC price [10]. Fourth, it has previously been observed that the intensity of care in the first month can also be a predictor of SMI [18].

In addition to personal suffering, a diagnosis of an SMI can also constitute severe consequences for service providers, such as GGZ Rivierduinen.

For one, previous studies have demonstrated that the group prioritized in its emergency service are patients with SMI [19]. The majority of these patients use emergency services, which are reserved for acute medical issues, such as first episode psychosis. SMI patients have higher rates of Emergency Department (ED) admissions and longer Length of Stay compared to other groups [3]. In the emergency service, the medical treatment expenses for each day of stay are very high, because they also include the expenses of the specialized practitioner, which puts pressure on the expenditure of mental healthcare centre. Evidence suggests that SMI patients have a higher DBC price in comparison to other disorders [12]. Besides, the length of stay and the more comorbidities a patient has, the more complex the treatment gets and that also influences the expenditure in consequence.

The DBC price will be implemented as the dependent variable to investigate if these risk factors also apply to the SMI population of this study, who are defined based on their high care requirement.

Over the last decade, a trend has emerged within GGZ Rivierduinen whereby a small portion of patients overuse and overload the resources and capacity of the centre [3][5]. Furthermore, GGZ Rivierduinen faces challenges with patients with SMI, that require and use a high amount of care and require help form different disciplines simultaneously. In GGZ Rivierduinen, the SMI population has not been studied, but is believed to be the cause of the overuse and overload of resources of the centre.

A key issue is that within GGZ Rivierduinen the SMI population is indicated differently than the epidemiological or Delespaul’s definition. This report aims to investigate the factors that contribute to the risk of belonging to SMI population and having a high total DBC amount.

The underlying aim of this investigation is to better prepare GGZ Rivierduinen through informing them of their SMI population’s characteristics and what their presence means for the service.

This aim has led to the development of the following central research question.

What factors influence the likelihood of a higher DBC price (in other words, belonging to the Severe Mental ill group) at GGZ Rivierduinen?

1.3. The Impact of SMI Patients on Healthcare Centres.

(14)

14

Following from the literature review, four main hypotheses were constructed that this study will examine: H1: The SMI population will have significantly higher frequency of contact and treatment hours than the rest of the GGZ Rivierduinen population in the first 30 days following admission.

H2: The SMI population will have more ED admissions than the rest of the GGZ Rivierduinen population.

H3: The SMI population has a lower chance of early discharge in comparison the rest of the GGZ Rivierduinen population.

H4: There is a positive relationship between the total DBC price (of SMI groups versus non-SMI group) by at least one of the following: gender, age, Emergency department admissions, living situation, education level, number of total comorbidities, department (to which department they belong to), type of diagnoses, LOS or treatment hours.

The following chapter discusses the research methodology, utilised analyses and model selection. It also provides an overview on independent risk factor and the dependent variable, the total DBC price. Chapter three describes the qualitative and quantitative analysis of the data and the findings of the analysis. In chapter four, the research findings and conclusions are outlined. Finally, in chapter five and six, the discussion, and the study’s implications, limitations and future research directions are presented, respectively.

1.5. Research approach

Figure 2|The research approach is outlined

For a better approach and understanding of the study population, the research included a qualitative part in addition to the collection and analysis of quantitative data.

(15)

Background

c

hapTer

2 2.1. Mental healthcare in the Netherlands

2.1.1. Structure of Mental Healthcare in the Netherlands

The insurance scheme in the Netherlands is based on a universal mandatory health insurance model. Under the national Health Insurance Law (Dutch: Zorg Verzekering Wet, ZVW), each citizen is obliged to have an individual private health insurance policy with an income-dependent contribution or subsidy [20]. This entitles the entire population to basic health insurance [21][22]. In addition, insured individuals are required to register with a single General Practitioner (GP), who coordinates and refers them for specialized care in hospitals, mental healthcare centres or other healthcare institutions as required [23]. GP’s belong to the primary healthcare, secondary care being hospitals and other places where you need a referral from a GP and tertiary being specialized healthcare centers (such as TBS clinics). Mental healthcare centres provide various forms of treatment: outpatient care; ambulant care; intensive care; and acute admission and emergency services [24].

Various healthcare providers including the Government and health insurance companies reimburse care expenses. Insurance companies can reimburse up to 100% of the care costs. The proportion of reimbursement that is provided depends on the type of basic insurance a client has at the time of the start of treatment [25].

The provided care depends on the required help and severity of the complaints For example, in order to receive specialised care, a GP referral and mental health assessment by a suitable professional are required [20].

Currently, mental healthcare centres are in the process of attempting to decrease the number of patients receiving treatment in specialised care [6]. This has resulted in a upwards trend in the number of patients treated in less specialised mental healthcare centres and a lowered likelihood of transfer to specialised care [26]. The insurance companies process divide treatments based on average treatments hours: those patients with less intensive and less severe mental disorder receive ambulatory care; while those with more severe disorder receive specialized care. The transition to specialized care occurs through utilization of the principal of shared decision making, in consultation of the patient [27].

Dutch health insurance companies cover each treatment on an individual basis, based on the required and requested care [Appendix B provided by KPMG][6][8]. This individualized process is conducted using the Diagnosis-Treatment Combination (Dutch: Diganose Behandeling Combinatie, DBC).

The DBC is an indicator which keeps track of the actions and costs that fit the provided care and assistance to a specialized care patient. The final DBC product is thus the sum of annual direct and indirect time that the patient has spent in treatment annually [25]. The total DBC price also includes the price of the days spent in an inpatient or outpatient clinic. If a client is admitted to emergency department (ED), a separate DBC with a fixed surcharge is opened [28]. Eventually, the Dutch Health Authority (Dutch: Nederlande Zorgautoriteit, NZA), who is in charge of the healthcare market in the Netherlands, specifies which treatments and DBC the Mental healthcare centres can declare to their insurer [29]. The health insurances, municipalities and care offices are also involved in negotiating with care providers on quality of care and price [15].

(16)

16 2.2. GGZ Rivierduinen

Geestelijke Gezondheidszorg (GGZ) Rivierduinen is one of the biggest mental healthcare centres in the Dutch province of Zuid-Holland. GGZ Rivierduinen provides mental health treatments to a spectrum of patients ranging from children, youth, adults and the elderly. The treatments provided also span a vast spectrum, from eating and feeding disorders, personality disorders, neurodevelopmental disorders, depressive disorders and psychotic disorders. GGZ Rivierduin’s main priority is to provide care for patients with complex illness profiles and who require a high amount of care. In order to provide the best possible care to its patients, multidisciplinary teams are involved in care planning, including psychiatrists, psychologists and social workers [30].

GGZ Rivierduinen collaborates with many other organisations such as GP practices, hospitals and other mental healthcare centres. This collaboration is crucial as effective sharing and communication of data, information and knowledge among stakeholders in the healthcare network is an essential factor for reducing healthcare errors and costs [31].

(17)

Methods

c

hapTer

3

The present study conducted a retrospective analysis on the patient data of the population of in- and outpatients within the clinics of GGZ Rivierduinen. The range of data points examined spanned the beginning of 2017 until the end of 2019. In addition, interviews with department directors and some researchers of the GGZ Rivierduinen were conducted.

In GGZ Rivierduinen, out of a total of N=31582 unique DBC’s examined, n=2039 were classified as having an SMI, over the period of 36 months. This period was chosen due to its recent nature, as well as completeness and availability of the data. Inclusion criteria were as follows:

* The patient was required to be older than 18 in 2017. * The patient had to reside in Specialized Care.

* The status description included completed and invoiced DBC’s of the patient. * The DBC file did not contain the Crisis DBC’s

The categorization of the population were ordered as SMI or not-SMI based on the level of care required, more specifically on the total DBC price. For a better understanding of the data, an Exploratory Data Analysis was conducted. The exploratory analysis concluded that a median DBC price should be utilised rather than a mean, as the data was not normally distributed. All participants with an Interquartile range (IQR) multiplied by 1.5 of the total were categorised as SMI. This way only the DBC outliers of GGZ Rivierduinen are scrutinized.

3.1. Study Sample

3.2. Operationalization of variables

A solid overview of the working process and information flow was required to get an insight of the GGZ Rivierduinen’s population. Interviews with department directors were performed to gain a better understanding of the study population and the flow of patients between different departments in the mental healthcare centres. The six directors who took part in the interview were brought into contact with the researcher by cluster manager van Harten. The interviews were conducted through the online telecommuting software Zoom. Following this, these interviews were summarized and used when writing the introduction.

(18)

18

Alongside a literature review, interviews with the department directors were performed to gain a better understanding of the study population and the flow of patients between different departments in GGZ Rivierduinen. One of the common points made by the directors was that the SMI population based on epidemiological characteristics is too limited of a definition. To be able to understand the impact the of the highly required care population, it was suggested to consider the whole of the SMI population in the research. This is because SMI, defined only based on epidemiological diagnosis, excludes patients with Personality disorder, Eating disorders and other disorders. “The last couple of years, number of patients with personality disorders are growing”, as mentioned by one of the 6 interviewee. Besides, one of the directors indicated that the SMI patients are unpredictable and, “usually need crisis service”. “Some of SMI populations flows in the mental healthcare centre from one department to the other”. However, the majority of this populations is treated in District ward (Dutch: Wijk teams).

3.2.2 Interviews

Datasets on individual DBCs, information on LOS and number of admissions in emergency service were available on the database ValueCare that is utilised by GGZ Rivierduinen. For further clinical data; like age, gender, education level and comorbidity, a medical investigator was approached. A Ms. Smith, helped to collect and to query data from GGZ Rivierduinen’s Electronic Patient Record (EPR), MyQuarant. Only DBC data from the claim years 2017 and 2018 were included. To optimise the data’s reliability, a second researcher, Ms. Bouma validated its use at each stage of data collection. This study was approved by the Ethics Committee of GGZ Rivierduinen.

Multivariate linear regressions were conducted to predict the impact of various variables on the DBC price for patients with SMI. The covariates for this regression were gender, age, education level, living situation, disorder/ diagnoses type and presence or absence of comorbidity. Before performing these analyses, the assumption for linearity were met for quantitative data. A p-value of 0.05 was used to assess the veracity of the null-hypotheses. In case of nonlinearity, a log-transformation was performed. This because, log-transformation transforms skewed data to approximately conform to normality [32].

In order to predict the impact of number of treatment and contact hours in the first 30 days on the SMI population, linear and logistic regression analyses were performed. In hypothesis 1, only DBC which were initialized in 2017 and 2018 were included. When investigating hypothesis 2, the relationship between admissions to Emergency Department (ED) and the SMI population, the same set of analyses (linear and logistic regression analyses), but here only DBC’s for ED admissions were included as variables.

Following this, in order to examine potential differences between SMI and the populations in their relative probabilities of early discharge, a Cox regression for survival analysis was conducted. This strategy was chosen as a Cox regression model is able to analyse multiple risk factors for survival. Finally, a neural network was built to examine the potential impact of various variables on a dependent variable, the total DBC price. Lasso and Garson’s algorithm function were used to evaluate relative variable importance in order to better understand the impact each variable had on total DBC price. Based on these analyses, appropriate variables were selected, and a set of analyses were performed (multivariate linear and multivariate logistic regression analyses).

The first set of analyses examined the impact of a few independent variables on the dependent variable and whether there was a significant difference between the SMI and non-SMI populations. The final hypothesis examined the impact and the importance of different variables on the outcome variable the total DBC price and SMI vs non-SMI.

The data analysis was performed in the coding software application RStudio.

3.3. Data collection

(19)

Results

c

hapTer

4 4.1 Preliminary Results

In order to provide a general overview of the sample population, a characteristics table is provided which outlines all the variables included in the study (Table 2). To predict the outcome of the total DBC price, several explanatory variables are used. After the descriptive results, this chapter is divided based on the four hypotheses that were presented in this study’s introduction.

The present chapter presents the results of the analyses that were performed on the dataset. As aforementioned, the SMI population was calculated by: 1.5* IQR+ Q3. The IQR of the total DBC price (IQR= 6064.8) gives us a value of € 9097. To calculate the outliers the value of 1.5 *IQR was added to the Q3(=7286.42) of the total DBC price, this gives the total value of €16383.65 The data was spread out in relation to the mean. All in all, the SMI population was determined to be those individuals who’s DBC price equalled or succeeded €16383.65, and the non-SMI population were those individuals whose DBC price dropped below this figure.

The results of the exploratory analysis showed that the SMI population at GGZ Rivierduinen makes up 45% of the total care expenditure of the healthcare centre, despite constituting merely 6% of the total population of the clinic (Table 1)(Figure 1). In table 1, the disorders represented in the SMI group are also shown as well as the total DBC price per disorder.

Table 1| The patient data of 2017 & 2018 is used, with all the closed and invoiced DBC.

GGZ Rivierduinen Number of _{population (%)} Total DBC price (%)

Total population 31582 203459493

Total Not-SMI population 29543 (94%) 111396012 (55%)

Total SMI population 2039 (6 %) 92063481 (45%)

SMI population

Bipolar disorders 243 (12%) 11002442

Depressive disorders 378 (19%) 16161701

Schizofrenia 620 (30%) 30891337

Anxiety disorders 169 (8%) 6981794

Feeding and Eating

disorders 108 (10%) 9264216

Personality disorders 169 (8%) 6010301

Pervasive 98 (5%) 5006460

Other disorders 154 (8%) 6745230

Table 1| The patient data of 2017 & 2018 is used, with all the closed and invoiced DBC.

(20)

20 4.2 Descriptive Results

Exploratory analyses demonstrated significant differences in the SMI and non-SMI groups in sixteen of the seventeen variables studied, only gender was found to be insignificant. Baseline characteristics of the patients are shown in table 2.

Significant differences could for instance be observed in the number of ED admissions for SMI population n=1901(93%) than other patients n=1863 (6%), see table 2. Furthermore, the median Length of Stay in the clinics for the SMI groups was Mdn=55 (IQR=75), in comparison with a median of Mdn= 15 (IQR=30) for the non-SMI group.

Significant differences were also found in types of diagnoses between SMI and non-SMI populations. For instance, n=620 (30%) of SMI patients held diagnoses of schizophrenic disorders, while a mere 14% (n=1863) of the non-SMI group were diagnosed as such. It furthermore shows that most patients with non-SMI come from the District Ward department n=809 (40%), while the majority of the non-SMI population originate from the policlinic (n=21010, 71%). The total DBC price also differed significantly between the SMI and non-SMI populations (Mdn=34118, IQR= 31633.3; and Mdn= 2503, IQR=3432.8, respectively) (Figure 3b, 3c)(Table 2). See full characteristics table in [Appendix C].

SMI n(%)/ mean (SD) Non- SMI n(%)/ mean _(SD) P value (n= 2039/ 31582) (n= 29543/ 31582) 49 (18.9) 48 (18.8) 48 (18.8) 0.009922 Anxiety 169 (8.3%) 3977 (13.5%) 4146 (13.0%) Bipolar 243 (11.9%) 2456 (8.3%) 2699 (8.5%) Depressive 378 (18.5%) 5880 (19.9%) 6258 (19.8%) Delirium 62 (3.0%) 1054 (3.6%) 1116 (3.5%) Feeding and Eating 208 (10.2%) 1198 (4.1%) 1406 (4.5%) Personality 169 (8.3%) 3387 (11.5%) 3556 (11.3%) Pervasive 98 (4.8%) 2803 (9.5%) 2901 (9.2%) Schizophrenic 620 (30.4%) 4149 (14.0%) 4769 (15.0%) Yes 1901(93.2%) 1863 (6.3%) 3764 (11.9%) < 2.2e-16 No 138 (6.8%) 27560 (93.7%) 27818 (88.1%) 55 (75) 15 (30) 31 (61) < 2.2e-16 1 (1.5) 1 (1.0) 1 (1) < 2.2e-16 Yes 1167 (57.2%) 4326 (14.6%) 5493 (17.4%) < 2.2e-16 No 872 (52.0%) 25217 (85.4%) 26089 (82.6%) 4.5 (0.98) 3.8 (1.0) 3.8 (1.0) < 2.2e-16 4 (3) 3 (2) 3 (2) 1.731e-11 Yes 1163 (57.0%) 14693 (52.9%) 15856(50%) 0.0003884 No 876 (43%) 13056 (47.1%) 13932 (50%)

Acute department 174 (8.5%) 612 (2.1%) 786 (2.5%) < 2.2e-16 District ward (Wijk) 809 (39.6%) 6298 (21.3%) 7107 (23.0%)

Eating disorder 205 (10.1%) 1321 (4.5%) 1526 (4.8%) Policlinic 739 (36.2%) 21010 (71.1%) 21749 (69.0%) Recovery ward (Herstel) 112 (5.5%) 302 (1.0%) 414 (0.0%)

158 (120.4) 22 (33.4) 24(40.4) < 2.2e-16 34118 (31633.3) 2503 (3432.8) 2541 ( 6064.8) < 2.2e-16

Table 2|(a) Data is expressed as mean ± one standard deviation. Probability determined using a two-tailed, independent samples t-test.

(b) Data is expressed as number within the sample who possess the characteristic. Probability determined using Chi square.

(c) For calculating the p-values of ordinal variables, Spearmans Rank correlations were used.

(d) Data is expressed as median + one interquartile range. Probability was determined using a Mann- Whitney U test.

< 2.2e-16

Departments n(%)

(c)

Treatment hours (hour) med (IQR) (d) Total DBC price (euros) med (IQR) (d) Length of stay (days) med (IQR) (d) ED admissions median(IQR) (d) ED admissions n(%)

(b)

Zorgvraag zwaarte mean(SD) (d) Total Comorbidities (n) med (IQR) (d) Comorbidity n(%)

(b)

Clinic admission n(%) (b)

Variable Total population _(31582)

Age (years) mean(SD) (a)

Diagnose type (disorder) n(%) (c)

Table 2|(a) Data is expressed as mean ± one standard deviation. Probability determined using a two-tailed, independent samples t-test.

(b) Data is expressed as number within the sample who possess the characteristic. Probability determined using Chi square.

(c) For calculating the p-values of ordinal variables, Spearmans Rank correlations were used.

(21)

Figure 3| The average price per unique costumer (APUC)

[a] the distribution of total DBC price of the whole population of GGZ Rivierduinen.

[b] The distribution of the total DBC price with non-SMI population it shift to an arbitrary looking distribution, with three different clusters.

[c] The median for the total DBC price of SMI population is much higher median= 34118, than the non-SMI population.

[a] [b] [c]

4.3 Number of treatment hours in the first thirty days

An ANOVA analysis revealed that a model including solely the number of treatment hours performed better than the model that added the frequency of contacts. A linear regression was therefore performed in order to predict total DBC price based on treatment hours of patients in the first 30 days of their treatment.

A significant positive regression was found (F(1, 15116) =2599.99, p < 0.001), with an R2 of 0.15 (Appendix D: Table 3.1; Table 3.2). In other words, a patients’ predicted total DBC price increases by 0.19% for every 1% increase in the treatment hours for the first 30 days. For example: for every 20% increase in the number of contacts, total DBC price increases by about (1.20 0.19 – 1) * 100 = 3.5 %. The total DBC price is equal to exp(7.38 + 0.19 *log(treatment hours) + e)(Equation 1), where treatment hours are measured in minutes (Appendix D: Table 3.3). The relation between the Total DBC price and the treatment hours in first 30 days is presented in Figure 4 (Appendix D).

A binary logistic regression was performed, to examine whether minutes spent in treatment (duration of contact) over the course of the first 30 days was a predictor for belonging to the SMI group or not. The dependent variable is coded “0” for the non-SMI and “1” for the SMI group. The SMI population had a median number of treatment hours of Mdn=908 (IQR=2316) while that of the non-SMI population was Mdn=145 (IQR=150)(Appendix D: Table 3.4).

4.3.1 Linear regression of treatment hours in first thirty days

4.3.2 Binary logistic regression of treatment hours in first thirty days

Equation 1|Linear Model Treatment Hours:

Linear equation: (ŷ)=β₀+β₁ x₁+ϵ

Log transformed equation: (ŷ)=e(β₀+β₁*log(x₁))+ϵ Fitted regression equation:

(22)

22

Table 3.5 demonstrating the number of treatment hours in minutes shows b = 0.0009, Wald = 887, p <0.001, OR =1.00 (95% CI: 1.000- 1.001), suggesting a significant difference in the SMI group in contrast with the non-SMI group (Appendix D: Table 3.5) (Equation 2). For example; for every two hours increase there would be (1.000944445708238)^120, this would be 1.12, hence and increase of 12% in odds ratio. This model proved significant, (χ2(1) = 1276.1, p <0.001), and explained around 20% of the variance (Appendix D:Table 3.6; Table 3.7). In other words, the treatment hours in first 30 days was a predictor of belonging to the SMI group, 93% of the cases were classified correctly when calculating the categorization strength, as shown by Table 3.8. Figure 5 shows that treatment hours of around 6000 minutes (100 hours) or more, in the first 30 days provides a roughly 90% probability of belonging to SMI group is (Figure 5).

Probability of belonging to SMI group given the independent variable treatment hours in the first 30 days.

Equation 2|Generalized linear model of treatment hours: Logistic equation: ln(P/(1-P))=β₀+β₁ x₁

Fitted regression equation:ln(P/(1-P))

ln(P/(1-P))=-3.06+0.0009* (x₁) [2]

Figure 5|Probability belonging to a group based on the treatment hours

4.4 Number of admissions to the Emergency Department

A linear regression was calculated in order to predict the total DBC price based on number of admission to the emergency department (ED). A significant regression equation was found (F(1, 31581) = 1691.07, p < 0.001), with an R2 of 0.05 (Appendix E: Table 4.1; Table 4.2). Patients predicted total DBC price increased by 1.13% for every 1% increase in the number of admissions to the ED. Example: For every 20% increase in the number of admissions to the ED, our total DBC price increases by about (1.20 1.13 – 1) * 100 = 23% percent. The total DBC price is equal to exp(7.90+ 1.13 *log(ED admissions) + e), where admissions to the ED is measured in numbers (Equation 3) (Appendix E: Table 4.3). The admission to the ED is a significant predictor of total DBC price. The relation between the Total DBC price and the ED admissions in numbers is presented in Figure 6 (Appendix E).

(23)

Equation 3|linear model ED admissions: Linear equation: (ŷ)=β₀+β₁ x₁+ϵ

Log transformed equation: (ŷ)=e(β₀+β₁*log(x₁)+ϵ) Fitted regression equation:

(ŷ) =e(7.90+1.13*log(x₁))+ϵ [3]

A binary logistic regression was performed, to examine whether number of admissions to the ED was a predictor for belonging to the SMI group. The dependent variable is coded “0” for non-SMI and “1” for SMI group. From the entirety of the SMI population, 57% (n=1167) were admitted to the ED at some point over the course of their treatment, contrasted with a mere 15% (n=4326) of the non-SMI population (Table2). This difference proved significant, (χ2(1)= 1431.5, p <0.001), and explained around 12% of the total variance (Appendix E :Table 4.4; Table 4.5). In Table 4.6, the number of admissions to the ED shows b = 0.76, Wald = 1306.7, p <0.001, OR =2.13 (95% CI: 2.04-2.22), suggesting a significant difference in the SMI group in comparison to the non-SMI group (Appendix E: Table 4.6) (Equation 4).

In summary, number of admissions to the ED was a predictor of belonging to the SMI group, 94% of the cases were classified correctly when calculating the categorization strength (Appendix E: Table 4.7). Figure 7 shows that the number of admissions to the ED of 7 admissions or more, gives around 90% probability of belonging to SMI group is (Figure 7).

4.4.2 Binary logistic regression of ED admissions

Equation 4|Generalized linear model of ED admissions: Logistic equation: ln(P/(1-P))=β₀+β₁ x₁

Fitted regression equation:

ln(P/(1-P))=-3.06+0.76 *(x₁) [4]

Figure 7|Probability belonging to a group based on the ED admissions

(24)

24 4.5 Probability of early discharge

A survival analysis was performed to test whether belonging to SMI had influence on the probability of early discharge. A total of n=30810 unique DBC’s were included in the study. 12 months after the start of 2017, around n=7443 (23%) of the DBC’s were ended and patients were discharged. After 24 months, n=10635 (34%) of the DBC’s were closed. Baseline characteristics of the patients are shown in table 5.1 (Appendix F).

Table 5.1 demonstrates a stark difference in the probability of discharge between SMI and non-SMI groups (p <0.0001). In the SMI population, n=264 individuals (13%) were discharged after 12 months, while n=7177 (24%) of the non-SMI population were discharged. The Kaplan-Meier survival analysis illustrates this difference (Figure 8). The Chi-Square test shows a high difference in observed and expected for SMI patients, this difference was significant (p <2e-16), as demonstrated in Table 5.2 (Appendix F). As this demonstrates that the difference between observed and expected results is highly unlikely due to chance, this suggests that other factors must be involved. The log-rank test for both groups is the same, suggesting no difference in survival curves (Appendix F: Table 5.2).

4.5.1 H3: probability of early discharge between SMI and non-SMI

Figure 8|Kaplan Meier curve for univariate SMI/ non-SMI

Early discharge survival during a follow-up period of 12 months. Kaplan-Meier estimates of the SMI group, (n=1999), non-SMI group (n=28812), p <2.2 e-16).

A Cox regression demonstrated a beta coefficient for SMI b = -0.80, which indicates that SMI patients have lower chance of discharge, compared to non-SMI patients. Thus, belonging to the SMI group reduces the hazard ratio by a factor of HR= 0.45 (95%CI: 0.40-0.50), or 33% (Appendix F: Table 5.3).

Finally, a multivariate cox regression was performed to examine which variables decrease the chances of an early discharge. A univariate Cox analysis was performed for various variables. From these, the significant variables were inserted into multivariate cox analyses. The results of univariate and multivariate Cox regression are given in table 8. Age and gender were coded into binary variables, with a division of male/female for gender, and age being divided as higher or lower than the mean patient age (μ=48 years).

(25)

The hazard ratio of for SMI stayed the same at HR= 0.45 (95% CI: 0.40 - 0.50) in univariate and multivariate cox regression HR=0.6 (95% CI, 0.51 - 0.74), and is still one of the most important variables in the multivariate cox regression (p < 0.001 ). The ward or department where a patient was treated proved to be the most important variable in the univariate cox regression (p < 2.2e -16), which stated that the risk for patients of not having an early discharge was HR= 0.04 (91% CI: 0.03- 0.04). The department where a patient was treated remained a significant predictor when performing the multivariate cox regression, (p < 2.2e-16), a risk of HR= 0.04 (95% CI: 0.03- 0.05) (Table5.4).

Another important variable that emerged from the univariate cox regression was the number of comorbid disorders (p < 2.2e-16), which showed that with an increase of 1 comorbid disorder, the risk of not having an early discharge is HR= 0.61 (95% CI: 0.59- 0.62), ignoring all other risk factors. This variable becomes somewhat less important in the multivariate analysis (p < 0.001), but still presents a significant increase in risk of not having an early discharge, namely HR= 0.74 (95% CI : 0.70- 0.78).

When including all the variables from the multivariate cox regression and looking at the Schoenfeld proportionality graph, the curves seems to remain roughly constant over time. It can therefore be concluded that the covariates are proportional enough(Appendix F: Figure 9).

Coef Hazard Ratio _{(95% CI)} P-value Coef Hazard Ratio _{(95% CI)} P-value

-0.80 0.45 (0.40-0.51) < 2.2e-16 -0.49 0.61 (0.51-0.74) < 0.001 *** -0.15 0.86 (0.82-0.90) < 2.2e-16 -0.04 0.96 (0.80-1.14) 0.64 -0.50 0.61 (0.59-0.62) < 2.2e-16 -0.30 0.74 (0.70-0.78) < 0.001 *** -0.00 1.00 (0.99-1.00) < 0.001 -0.00 1.00 (1.00-1.00) < 0.001 *** -0.33 0.72 (0.70-0.73) < 2.2e-16 0.07 1.07 (0.98-1.17) 0.13 0.02 1.02 (0.97-1.07) 0.394 Recovery ward (herstel) -2.62 0.07 (0.06-0.09) < 2.2e-16 -3.12 0.04 (0.02-0.10) < 0.001 ***

Poli -1.83 0.16 (0.14-0.18) < 2.2e-16 -2.49 0.08 (0.07-0.10) <2.2e-16 ***

District ward

(wijk) -3.31 0.04 (0.03-0.04) < 2.2e-16 -3.31 0.04 (0.03-0.05) <2.2e-16 *** Eating

disorder -1.61 0.20 (0.17-0.23) < 2.2e-16 -1.89 0.15 (0.11-0.21) <2.2e-16 *** Table 5.4| The univariate and multivariate results of the risk factors

Total comorbidities Length Of Stay (n) Zorgvraagzwaarte Gender (Female) Department Age (old >48) Univariate Multivariate Variables SMI

(26)

26 4.6 Influence of the risk factors on the total DBC price

Figure 10 shows how that weights of each input variable is calculated and added to the weight for hidden layers, and this way the value of the output layer can be calculated.

A neural network (NN) analysis was performed to compute an estimate total DBC price (dependent variable) (tDBCp), based on various risk factors (independent variables) including: total treatment hours (TH); LOS in clinic (n) and/or to ED (ED); which department the patients belongs to (such as the Acute, Eating, Recovery, Poli and District Wards); and number of comorbidities (TC) (Figure11). The training process of NN needed 70 steps until all absolute partial derivatives of the error function

were smaller than 0.0096 (the default threshold). The estimated weights after performing the NN ranged from 5.14 to 8.68. For instance, the intercepts of the first hidden layer are 6.82 and 7.88 and the nine weights leading to the first hidden neuron are estimated as 7.60, 5.14, 7.26, 8.68, 6.55, 6.45, 8.24, 6.83 and 7.52 for the covariates ED admissions, Treatment hours, number of total comorbidities, LOS, and belonging to department Acute, Eating, Recovery, Policlinic and District Wards., respectively (Appendix G: Table 6). These weight contribute to calculate an estimate of the total DBC price based on the variables in the input layer (See Appendix G: Equation 5).

In Figure 11, the plot reflects the structure of the network, it includes by default the trained synaptic weights, all intercepts, it also includes some information about the training process like the overall error and the number of steps needed to converge.

4.6.1 Neural Network for prediction of total DBC price

The neural network exists of input independent variables, hidden layers and the dependent variable in the output layer.

Figure 10| Illustration of Neural network

Figure 11| Neural network of the risk factors

Plot of a trained neural network including trained synaptic weights and basic information about the training process.

(27)

To determine each variable’s relative importance in the NN model, Garsons’s algorithm and Olden’s connection weights algorithm analyses were performed (Figure 12)(Appendix G: Figures 15 and 16). The importance values assigned to each variable are also in units based on the summed product of the connection weights, whereas Garson returns importance scaled from 0 to 1. This shows that according to Garson’s algorithm the LOS, ED admissions and belonging to the ward District are most important factors of the NN model and influence the dependent variable (Figure 12).

Figure 12| Garson’s relative importance algorithm

To elaborate on the NN model and to predict the total DBC price, multiple linear regression was performed based on: the number of admissions to Emergency Department, treatment hours , total number of comorbidities , Length of Stay in clinic to as well as the department the patients belong to. A significant regression equation was found (F(8,31308)=5395.37, p< 0.001, with R2 of 0.88 (Appendix G: Figure 7.1; Table 7.2).

Patients predicted total DBC price was increased by 0.17% for every 1% increase of LOS, and it increased by 1.04% for every 1% increase in treatment hours. This while the total DBC price for instance increased by 0.78% if the patient belonged to the District ward (Appendix G: Table 7.3).

Patients’ predicted total DBC price is equal to exp(5.36+ 0.17*log(LOS) – 0.11 * log(TC) + 1.04 *log(TH) – 0.06 *log(ED) – 1.14 *log(Eating)+ 0.26 *log(Recovery) – 0.72 * log(Poli) – 0.78* log (District ward) (Appendix G: Table 7.3).

For example: For a patient with 55 days LOS, 3 TC, 600 TH, 2 ED admissions and who belongs to the District ward, the total DBC price increases by about: exp(5.36+ 0.17*log(55) – 0.11 * log(3) + 1.04 *log(6000) – 0.06 *log(2) – 1.14 *log(0)+ 0.26 *log(0) – 0.72 * log(0) – 0.78* log (1)= €13556.13 (Equation 6).

All the variables included in this model were significant predictors of the total DBC price. The relation between the Total DBC price and all these variables are presented in Figure 15 (Appendix G: Figure 15) .

4.6.2 Linear model of the risk factors on the total DBC price

Equation 6| linear model chosen variables: Linear equation: (ŷ) =β₀+β₁x₁+...+β_px_p+ϵ

Log transformed equation: (ŷ) =e(β₀+β₁*log(x₁)+...+β_p*log(x_p)+ϵ) Fitted regression equation:

(ŷ) =e(5.36+ 0.17* log(x₁) – 0.11 * log(x₂) + 1.04 * log(x₃) – 0.06 * log(x₄) – 1.14 * log(x₅)+ 0.26 *

(28)

28

A binary logistic regression was performed, to examine whether various variables like: the number of admissions to ED, total number of comorbidities , LOS in clinic to and the department the patients belong to were predictors for belonging to the SMI group or not. After performing ANOVA, treatment hours (minutes) was excluded from this model. This model proved significant, (χ2(7)= 12396, p <0.001), and explained around 99% of the variance (Appendix G: Table 7.4 ; Table 7.5 ).

In Table 7.6, the number of admissions to the ED shows b = 0.071, Wald = 6.015, p <0.001, OR =1.074 (95% CI: 1.016- 1.138), suggesting a significant difference in the SMI group in comparison to the non-SMI group (Table 7.6). And belonging to the District ward increases the odds of belonging to SMI population 3 times more, b = 1.178, Wald = 63.77, p <0.001, OR =3.248 (95% CI: 2.391- 4.444). This while per unit increase in TC the odds of belonging to SMI decreases by 0.93, with b= -0.068, Wald= 12.03, p <0.001, OR =0.93 (95% CI: 0.899- 0.971) (Appendix G: Table 7.6). The only not significant variable was belonging to Eating & feeding department (Appendix G: Table 7.6). The logit model the response variable is log odds: ln(P/(1-P))=-1.289+0.011*(LOS)-0.068*(TC)+0.07*(ED)+2.079*( Recovery)+0.860*(Poli)+1.178*(District ward) (Equation 7).

In summary, LOS, number of admissions to the ED and belonging to a specific department were predictors of belonging to the SMI group, 59 % of the cases were classified correctly when calculating the categorization strength (Appendix G: Table 7.7).

4.6.3 Binary logistic regression of the risk factors on the total DBC price

Equation 7|Generalized linear model chosen variables: Logistic equation: ln(P/(1–P))=β₀+β₁ x₁+...+β_k x_k

Fitted regression equation:

(29)

Discussion

c

hapTer

5

This study aimed to investigate the following research question: What factors influence the likelihood of a higher DBC price (or belonging to the Severe Mental Ill group) at GGZ Rivierduinen? Four hypotheses were developed, investigating the relationship between the SMI population, DBC price and a) frequency of contact and treatment hours in first 30 days; b) Emergency Department admissions; c) early discharge and d) demographic characteristics. Significant relationships were found for all four hypotheses, thus demonstrating that the SMI population differed substantially from the non-SMI population.

With respect to the first hypothesis, it was found that treatment hours in the first 30 days was a significant predictor of the total DBC price. In other words, the more treatment hours a patient received, the higher the total DBC price rose as a consequence. This same result was found for treatment hours in SMI compared to non-SMI groups: the higher the number of treatment hours, the higher chance the individual belonged to the SMI group. In other words, the SMI population had more treatment hours in first 30 days in comparison to the non-SMI group. These findings align with those found in a literature review conducted by Ten Have and their colleagues (2012), who also found the intensity of care in the first month to be a predictor of SMI.

However, we must be careful in interpreting the results of the present study. For instance, this study was unable to make a distinction between DBC’s with missing data on the number of treatment hours, and those patients who had zero treatment hours in the first thirty days of admission.

In other words, DBC’s without information about the number of treatment hours in first 30 days were also treated as having no treatment hours in first 30 days. This might have influenced the results of the performed regression. If all patients had information on their number of treatment hours in first 30 days, an even higher difference could have been observed between the SMI and non-SMI in terms of their treatment hours in first 30 days, but also the opposite could have been detected.

Evidence was furthermore found for the second hypothesis: the number of Emergency Department Admissions had a positive relationship with the total DBC price, and thus belonging to the SMI group. Results indicated that the odds of being admitted to the Emergency department was two times higher when a patient belonged to the SMI group. Thus, the more ED admissions a patient has, the higher the chance that that patient would belong to SMI population.

Saurman and their colleagues (2014) as well as Thomas and their colleagues (2018) have also reported that the SMI population they studied had considerably more ED admissions than other patients.

These findings provide further evidence in favor of Delespaul’s definition of SMI’s, as his definition indicates that Emergency Department admissions are a valid predictor of SMI population belonging. However, a point that must not be overlooked is the fact that the Emergency Department is the only department open during weekends and non-working hours. This means that patients could be admitted to the ED despite the fact they may require a lower level of care, simply because no other department is open at the time. This could result in their DBC price being driven up, despite the fact that they do not warrant belonging to the SMI group, and therefore have biased the results of this study.

(30)

30

Findings demonstrated a significant difference in the probability of early discharge between the SMI and non-SMI population, confirming the predictions of the third hypothesis. Individuals in the SMI-population were found to be in care longer than those in the non-SMI population. This means that the SMI population is more likely to be treated for longer periods of time and utilise more services than the rest of the GGZ Rivierduinen population. However, it is important to consider early discharge alongside other variables, such as ED admissions and number of comorbidities, which were found to also contribute to the delay in discharge in the SMI population. Of course, if a patient requires more complex treatment and more intensive treatment, this lowers the probability of that patient being discharged early and drive up the DBC price, thus heightening the chances of belonging to the SMI group.

It is also possible however that these results were skewed due to the limitations posed by only examining patient discharge between the years of 2017 and 2018. This is because a patient discharged in 2018 might have been readmitted the following year, without this study being able to observe this. In addition, if a patient was admitted in late 2018, and got discharged early 2019, the patient would not be considered as discharge early due to lack of follow up.

The fourth hypothesis demonstrated various outcomes: in contrast to previous literature, no relationship was found between the total DBC price and following variables: age, gender, education level, living situation and diagnosis type.

Belonging to an older age category was however found to be predictor of SMI in a study done by Hunt and their colleagues (2019) and Kennedy and their colleagues (2019), while this same result was not found in the present study. This could be due the fact that the majority of the GGZ Rivierduinen population is below sixty years old ; had there been more patients above the age of 60, findings by Hunt and Kennedy could have potentially been replicated in the present study.

According to Hunt and their colleagues (2019), female gender was found to be a predictor of belonging to the SMI group, while this study failed to find such a connection. The equivalence in results of men and women in this study might have been caused by the frequency of women using alternative caretakers instead of mental health organisations in comparison to men, as claimed by Schuch and his colleagues[33].

Findings from the present study also contradict those of Delespaul (2001) in socio-economic status. Delespaul found differences in socio-economic status of patients with an SMI classification compared to those who did not. These differences may not have been established in the present study due to a skew in missing data. 31% of the data for the SMI population was missing, while this was the case for only 19% of the non-SMI population. Contrary to expectations, the diagnosis type was also found to be an insignificant predictor of the total DBC price. This suggest that the healthcare centers can disregard the diagnosis type of patients when investigating the amount of required care, or when exploring the total DBC price. However, it is still an important factor for determining the SMI population and when deciding for instance on treatment type.

Findings demonstrated that the longer the Length of Stay, the higher chance that a patient belonged to the SMI group and higher the total DBC price. In addition, the department that a patient belonged to also changed the likelihood of belonging to the SMI group. For instance, if a patient belonged to Recovery ward or District ward, they were more likely to belong to the SMI group, than if they belonged to the Policlinic ward.

Contrary to expectations, it was shown that the SMI group had fewer comorbidities than their non-SMI counterparts. Furthermore, a negative relationship was detected between the total number of comorbidities and the total DBC price. This result could be due to the fact that the number of comorbidities from the DBC’s in 2017 and before were measured and recorded based on the DSM-4 classification system. In contrast, the DBC’s from 2018 were recorded based on DSM-5 criteria. These two classification systems were recorded and stored differently in the electronic patient record of GGZ Rivierduinen. This resulted in lower total number of comorbidities for patients whom there DBC started in 2018.

All the significant demographic variables were included in The Neural Network model, all the variables demonstrated a positive relationship with the total DBC price.

(31)

5.1. Implications

5.2. Limitations

This study confirms some of the findings in previous literature, such as that Emergency Department admissions and Length of Stay are predictors of the SMI population. What is surprising is that belonging to a specific department was also found to be a significant predictor of the SMI population at GGZ Rivierduinen. Other healthcare centers should therefore be aware of these differences in their departments. It was also demonstrated that the treatment hours in first 30 days was a predictor of the SMI population. These findings mean, that it would be possible to predict which patients will on the longer run belong to SMI population, by looking at their treatment hours in the first 30 days after admission. These results indicate that the SMI groups is complex and requires significantly more care than the rest of population of GGZ Rvierduinen.

This study demonstrated that for the purposes of calculating healthcare costs, there is little use for the epidemiological definition of SMI’s, as diagnosis classification did not correlate with total DBC cost. In other words, in terms of healthcare burden, the type of diagnosis a patient had did not matter much, while other factors were far more important.

This report included all the patients who had a complex profile of illness and required a high amount of care. It helped us to understand the need and complexity of this group and it did not exclude patients based only on epidemiological characteristics such as psychosis.

Prior to this study, evidence of the existence of an SMI population at GGZ Rivierduinen was purely anecdotal. This study was the first to provide quantitative evidence of their existence, their function within the healthcare center as well as their healthcare needs.

These outcomes, in particular the Neural Network, could also be used for developing a simulation to predict SMI population and this way to provide and prepare better care for this group.

For instance, if data for January 2021 is available on the number of Emergency Department admissions, the average length of stay, the number of treatment hours, then the neural network model can be used to predict DBC price per patient, and therefore predict the expenditure of GGZ Rivierduinen.

It is also possible to utilise these results in order to develop targeted interventions aimed to reduce Emergency Departments admissions or to help accelerate discharge process. Targeted intervention, such as hiring or training the staff of the Emergency Department for more complex patients or organizing more multidisciplinary consultations when accounted with SMI population, can be arranged.

The generalizability of these results is subject to certain limitations. For instance, in this study, the unique DBCs are taken into account instead of unique individuals. Maybe if the research was done on patient level, we might have accounted and obtained more specific risk factors. Besides, the diagnosis type was excluded in the multivariate linear model due to insignificance, while in literature it is shown as one of the most important predictors of SMI. This indicates that maybe a difference in SMI population based on epidemiological definition should have been investigated alongside the SMI definition based on quantity of care. This would have given a better insight on what factors have an impact on SMI population, with different definitions.

In many literature reviews addictions to drugs and mental retardation were implied as important predictors for belonging to SMI, however these two variables were not included in the study. This is because these two variables could not be obtained from the Electronical Patient Record, due to the way they are registered or they are not registered at all. By adding these two possible risk factors the epidemiological characteristics of SMI population could be indicated, explained, and substantiated.

Furthermore, there was also missing data in various variables such as education level, living situation, which might have affected the results. The data might have ensured that there is or there is no difference found for these factors in SMI group versus non-SMI.

Despite its limitations, the study definitely has usable results on a large population sample, across a range of departments, diagnoses, and years. This adds to our understanding of the role of SMI population on expenditure in GGZ Rivierduinen and therefore, this is a great first step.