Hospitalization-associated disability: using statistical learning to reduce the dimensionality of data and promote analysis

(1)

Hospitalization-associated disability:

using statistical learning to reduce

the dimensionality of data and

promote analysis

STEVEN W.J. NIJMAN

(2)

(3)

2 Student S.W.J. Nijman, MSc Student number: 10885846 Email: s.w.nijman@amc.uva.nl Mentors M.C. Schut, PhD Email: m.c.schut@amc.nl J.J. Aarden, MSc Email: j.j.aarden@hva.nl

Department of Medical Informatics Department of Geriatrics

Academic Medical Centre University of Amsterdam

Tutor

J.H. Leopold, PhD Faculty of Medicine

Department of Medical Informatics, AMC-UvA Email: j.h.leopold@amc.nl

Location of Scientific Research Project

Department of Medical Informatics

University of Amsterdam – Faculty of Medicine Meibergdreef 3

1055 PE Amsterdam The Netherlands

Practice teaching period

(4)

(5)

4

Preface

This thesis is a description of the scientific research project (SRP) I did for the past 8 months at the department of Medical Informatics (KIK) at the Academic Medical Centre (AMC) in Amsterdam. What initially started as a simple question about activity trackers and a stack of paper with too many variables to count, turned into a great exploration about statistic possibility and data science. It has been most rewarding and I am eager to continue pursuing a career in a similar discipline.

I would like to extend my thanks towards my supervisor Martijn Schut and my tutor Hemmik Leopold for their warm welcome and refreshing collaboration. With your enthusiasm and ideas I was continually motivated to strive for more and learned a lot about a field I now consider my future. It has been a remarkably pleasant feeling to work with someone rather than for someone and I can’t thank you enough for that experience. I look very much forward to the project we will do together this summer.

Next, I would like to thank all other people involved in the project that extended a helping hand in discovering how to go about my work. Much gratitude to the geriatrics department who set up this research and enabled me to do this project.

Lastly, I would like to give a special thanks to my parents, brother and girlfriend for their continuing belief in me. Thank you.

(6)

5

Chapter 1

(12)

9

1.1 Introduction

The healthcare system is being influenced in a major way by developments in data science and Big Data research. The advent of technology paves the way to larger scores of data and significant cost reduction for most medicine practices [1, 2]. These concepts are gaining a solid foothold in how we practice medicine and conduct clinical research, as can be observed with the widespread implementation and utilization of the electronic health record (EHR) [1, 3]. As such it is highly beneficial to be known with the potential uses and risks for these technologies.

Data science can be described as ‘a set of fundamental principles that support and guide the

principled extraction of information and knowledge from data’ [4]. A form of data science is

statistical learning, which utilizes machine learning. Big data is seen as a large volume of data, often with such high complexity and variety that regular statistical software is unable to process it at an acceptable pace [1, 4]

This thesis describes two related studies which were performed during the period of November ’17 till June ’18. Each study focusses on a different application of statistical learning to the concept of hospitalization-associated disability. The overall objective of both studies was to evaluate the viability of statistical learning methodologies on clinical research data such as the one the Hospital-ADL study provides and to provide a provisional first analysis of this data to Geriatrics department of the Academic Medical Centre (AMC) in Amsterdam [5].

1.2 Outline of this thesis and description of studies

The next chapter of this thesis will outline background information on hospitalization-associated disability and statistical learning. This is primarily to ensure the reader has a full understanding of what the methods entail and why the posed questions are tried to be answered. For hospitalization-associated disability we will present a general introduction, a brief history, acute hazards, common predictors and what the current state of research is. For statistical learning we will introduce techniques for dimension reduction, the concept of correlation matrices and rotation, linear regression, information criteria and predictive outcome measures.

The first study presented in chapter 3 of this thesis, describes a detailed analysis of the variables introduced in the hospital-ADL study. The study evaluated which variables would be most important using statistical learning in form of the data reduction technique called factor analysis. Outcomes of the factor analysis and adequacy measures are provided. We used statistical learning principles to gain insight in what components provided and what variables considered were deemed most important in comparison to each other. It is especially interesting to see if the predictors introduced by literature are equivalent to the variables that cause most variation in this analysis.

The second study presented in chapter 4 of this thesis, describes the development of a predictive model concerning for the Katz index of Independence in Activities of Daily Living. For the purpose of reducing the amount of dimensions we again used factor analysis. This time the Katz score was not included in the list of provided variables and instead was used as the outcome variable. We then used regression using information criteria such as BIC and AIC to find the best fitting model. Subsequently the predictive power of the model was evaluated and reported.

(13)

10

Chapter 2

(14)

11

2.1 Hospitalization associated disability

2.1.1 Introduction

Hospitalization-associated disability (HAD) is a complex problem that has received public interest for many years. HAD is defined as a patient incurring new disabilities, specifically concerning a patient’s ability to perform activities of daily living (ADLs), because of their hospitalization. It quickens the decline of functional status in those unable to walk, wash or dress well [6]. Unfortunately this occurs often, especially in the elderly [7-10]. Its prevalence is immense as 30 to 60 per cent of acutely hospitalized elderly patients run the risk of accruing functional loss during their stay [11]. The efforts to conceptualize a trajectory of functional decline during and after stay has so far been difficult as it was not observed to be entirely predictable [12, 13]. Nevertheless, functional decline can be observed as an important predictor for discharge outcomes [14-16]. An elderly patient, after acquiring a new disability during stay, receives prognosis similar to that reported for disastrous conditions such as a hip fracture or stroke and only has a thirty per cent chance of restoring their functional performance [14]. When given time (Boyd’s study used a 2-year follow up) it has been observed that patients can recover close to their pre-admission level of ADL-functioning [17]. The first three months are most indicative of how successful the patient’s recovery will be [5, 17]. Proper treatment and the possibility of nursing care are essential for recovering to acceptable levels of functional status [18].

2.1.2 A brief history

The staff of the Benjamin Rose Hospital in Cleveland, Ohio, have long focussed their efforts on obtaining accurate measurements methods for functional loss. In 1959, as a part of this effort, they defined an Index of Independence in Activities of Daily Living (ADL), which proved to be much more accurate than other measurement methodologies [19]. The original study of the measurement methodology initially only observed elderly diagnosed with a hip fracture. In 1963, the index broadened its validity by demonstrating wide applicability [20]. Developed as a way of studying the effectiveness of treatment in elderly patients, by looking at their ability to perform basic activities of daily living, the index was created because improved measures of function were missing [20]. According to Rosin, in 1964, it was well known that functional disability was an important influencer for patient discharge outcomes [21]. Later, it was suggested that whilst treating with the best intent, care institutions caused geriatric patients to be put into settings of general inactivity, which worsened their functional ability [22]. This was also established in a later study which showed that hospitalization may have a negative influence on the functional status of an elderly patient, even if the surgery or treatment were to be successful [23].

2.1.3 Acute hazards

Whilst a period of 3 months to two years for near-full recovery may seem acceptable the acute hazards of HAD are severe and thus prevention is often preferred to recovery [11, 24-26]. Mudge et al. proposes that it may be more beneficial to advocate for early functional recovery as, in their study, the rate of functional decline pre-hospital was higher than the in-hospital rate [27]. Mudge et al. suggests that their conclusions are limited as the prevalence of HAD varies between health care systems. Therefore early functional recovery will not be discussed further in this study. An example

(15)

12

of an acute hazard can be observed as higher 3-month mortality when an elderly patient is unable to regain their baseline functional status after hospitalization [13]. Other observed hazards include, but are not limited to, depression, decreased mobility, decreased quality of life, decreased functional independence, cognitive impairment and a higher risk of nursing home admission [9, 10, 12, 23, 28-30].

2.1.4 Common predictors

HAD is expected to be influenced by a large variety of factors and the mechanisms underlying HAD are elusive [5, 12]. Some common predictors for functional decline that researchers agree on are age, malnutrition, muscle strength, frailty, pre-admission need for assistance, length of stay, cardiovascular disease, dementia, albumin levels, depression and low mobility levels [11, 14, 15, 17, 18, 24, 25, 30-33]. A low level of education was found to be a factor in several studies, however Boyd et al. argues that instead a higher education caused functional decline more often [11, 17, 24].

2.1.5 Current work

Using these common factors, and other findings, studies have attempted to create an environment where the elderly are subject to better care [32, 34-36]. Unfortunately many have shown to be ineffective or ill-attuned to the wishes of the patient [37]. Other studies have stressed the importance of future work, specifically in prevention, screening, including functional information, HAD-focused care model adoption, understanding of rehabilitative services, and understanding of patient’s functional trajectory [7-9, 11, 16, 17, 24, 31, 38]. Ettinger et al. emphasizes the importance of hospitals and health systems working together to test their projects and report the results publicly [9]. Unfortunately, for the studies involved, the enormous methodological variability between studies makes meta-analysis or comparison difficult [39].

2.1.6 The Hospital-ADL study

The Hospital-ADL study developed by the geriatrics department at the AMC tries to combine several methodologies under one umbrella [5]. As such the clinical informatics department of the AMC was asked to provide its statistical knowledge and perform statistical learning methodologies with the provided data of the Hospital-ADL research protocol.

(16)

13

2.2 Statistical learning

2.2.1 Introduction

Understanding and modelling complex data sets is central to statistical learning methodologies. It is a developing field in statistics and machine learning that, especially as ‘Big Data’ is expanding, has much value in many scientific areas as well as other disciplines. Statistical learning is often separated in supervised and unsupervised tools. Statistical learning with unsupervised data means that there is no outcome variable. Whilst no prediction can be made, relationships and structures between variables can still be explored. Supervised statistical learning often includes predicting an outcome variable, given a set of one or more input variables. The type of outcome variable motivates what kind of problem needs to be solved in supervised statistical learning. This can be either a regression problem, when dealing with a continuous range of values, or a classification problem, when dealing with a discrete set of labels.

2.2.2 Dimension reduction

Reducing the dimensionality of data is common in statistical learning. Often used to decrease computational cost or ensure the chance of overfitting is minimal [40].Overfitting means the model will perform well on the training data, but poor on the testing data. Think of over-fitting as a situation where a model commits the training data to memory, rather than learning to know why the data behaves that way, which will prevent it from being able to be generalized to test data, which the model has never seen. Two common methodologies in dimension reduction are principal components analysis and exploratory factor analysis.

Principal components analysis

Principal components analysis (PCA) has a multitude of different uses [41]. In general it is used to discover a set of features, or principal components, from a larger group of variables, often as a way of reducing the dimensionality of data. PCA is also usable with unsupervised data as a part of exploratory data analysis. The main idea represented by PCA is to summarize a large dataset with a smaller number of principal components which collectively explain most of the variability of the original set [41]. The first principal component would show the direction of the data that varies most, with each next principal component explaining the next direction that varies most. The variables used for PCA need to be continuous, although using dummy variables makes it possible to include a set of discrete variables. A dummy variable means the discrete value is converted from a string or category to a 1 or 0 to indicate whether or not some effect is given. Statistically speaking principal components represent a combination of random variables that show the maximum variance. These variables do not need to be related and the components do not need to be interpretable.

PCA can theoretically have n – 1 distinct principal components (where n is the number of observations [41]. In practice the first few principal components are often selected given they visualize the data or allow us to interpret the results well. The number of principal components to retain is done via a multitude of methods, which will not be covered in this study.

(17)

14

Exploratory factor analysis

Exploratory factor analysis (EFA), or common factor analysis (CFA), is used to explore the latent underlying relationships within a group of variables [40]. This results in a smaller number of factors which can be used to represent relationships among sets of interdependent variables. The factors often represent concepts not directly measured in the variables themselves. These are called latent factors. EFA is in many ways comparable to PCA, but is still distinctly different [40]. The difference is that in PCA the combinations of variables is random, given the maximum variance they may represent. EFA assumes that each individual variable scores are a result of the underlying latent concept it is trying to represent. Where PCA is not constructed for interpretation, EFA is.

A more distinct different lies in the amount of variance included in the solution. Where PCA includes all three types of variance, EFA only uses one type of variance: common, or shared, variance. The three types of variance are common variance, specific variance and error variance [40]. Common variance is the shared variability of a variable with other variables, specific variance is the variability explained by the unique components of a variable and error variance generally indicates the measurement error and unreliability of the variable. It is suggested that because all variance is included in PCA the estimates provided give inflated results [42]. This discussion lies outside the scope of this study and will therefore not be discussed.

In similar fashion as principal components analysis, the amount of factors can be high. Whilst using a lot of factors will explain all variance in the data it is infeasible as each factor is supposed to represent a latent factor. In that sense it differs from principal components analysis as it aims to have interpretable results. As each factor represents a latent concept explained by a combination of variables, often not directly recorded by the study, there is a limit to the amount of relevant concepts the data holds. Therefor a stopping rule needs to be put in place, to ensure the amount of factors is relevant given the provided variables. Normally the amount of factors to be retained is calculated by a specified stopping rule in a factor retention methodology.

Horn’s parallel analysis has garnered much support over the years as the best empirical method for factor retention in factor analysis [43]. It is an adaptation of the Kaiser criterion, the widely-used standard for factor retention and similar to the broken-stick criterion in that it uses random data [44]. It is suggested that whilst the broken stick model is not necessarily the best, an adaptation or consensus between stopping rules will work best [45].

2.2.3 Covariance and correlation matrix

A correlation matrix is a normalized covariance matrix, with a mean of 0 and a standard deviation of 1.0, giving the correlation coefficients between all pairs of variables in a data set [46]. Whilst covariance matrices have their place in statistics, correlation matrices are used most often. One advantage with a correlation matrix is that it normalizes variables, enabling them to be used together in analysis, regardless of their unit types.

A correlation matrix is presented in a symmetrical arrangement providing systematic measures of the relationship among a large number of variables. Most often Pearson’s correlation is used, representing the linear association between variables X and Y. It is important to keep in mind that even though correlation may be established, causation is not automatically proven.

(18)

15

Multivariate statistical techniques, such as PCA and EFA, often use a correlation matrix as its starting point. EFA generally looks for variables that have a high positive correlation with a group of other variables, whilst having a high negative correlation with those variables outside of that specific group.

When creating a correlation matrix it is assumed all variables are continuous. When working with mixed data, including dichotomous and ordinal variables, Pearson’s correlations are not suitable [47]. Dichotomous variables require tetrachoric correlations to be able to legislate themselves in a matrix containing continuous variables. Between continuous and ordinal data polyserial correlations are necessary to establish correlation coefficients. Correlations of ordinal data are calculated using polychoric correlations. A correlation matrix where all these different types of correlations are combined is called a heterogeneous correlation matrix.

2.2.4 Rotation

When factor analysis was conceived the results were often difficult to interpret, requiring visual aids to subjectively decide which factors were apparent in the results [48]. Soon after the conception of factor analysis a technique called rotation was developed in order to improve the initial interpretability of the results [48]. The term “rotation” was picked as the axes were essentially rotated so that the combinations of variables grouped for each factor would fall in close proximity to them (figure 1).

Figure 1: Orthogonal and Oblique rotation [49]

The two types of rotation developed are orthogonal and oblique. Each has a number of different algorithms that are meant for slightly different data sets, whilst still following the assumptions of either orthogonal or oblique rotation. The most popular orthogonal rotation algorithm is generally Varimax whilst the most popular oblique rotation is Promax [50, 51]. Discussing the algorithm in detail falls outside the scope of this study and will therefore not be discussed.

(19)

16

The key difference between the two types of rotations is that orthogonal rotations assumes factors don’t correlate, whilst oblique rotations assume they do [48]. Generally in the social and medical sciences it is recommended to use oblique rotation as it is difficult to prove factors are not correlating in some way with each other [48].

2.2.5 Linear regression

Regression is one of the simpler supervised methodologies used for predicting a quantitative outcome variable. In statistical learning linear regression is seen as a good starting point for further analysis [41]. Linear regression is still actively used to this day, especially for finding causal relationships between two variables. It also is used to infer how strong this relationship is. Simple linear regression is viable for predicting an outcome variable using a single predictor. Having more data, which in practice happens more often than not, requires a more suitable method. As fitting separate linear regressions for each relationship is not feasible [41]. Fortunately it is possible to extend the model, given that it is linear, so that each predictor has a separate slope coefficient, accommodating multiple predictor variables [41]. This is called multiple linear regression.

A regression does not necessarily use all variables given in a dataset. It makes a selection of a combination of variables which best characterizes the dataset and provides the most accurate prediction whilst promoting parsimony. A parsimonious model means it is simple, interpretable and generally has a low number of predicting variables [52].

2.2.6 Information criteria / selecting a model

Basing a viable model on the information given by a number of finite variables is difficult, which is why Hirotugu Akaike developed the Akaike Information Criterion (AIC) [53]. Think of the AIC as a number that is helpful for comparing regression models as it includes measures of how well the model fits the data as well as how complex the model is. For linear regression some selection criteria, such as Mallows’ Cp, are asymptomatically equivalent to AIC, meaning they eventually become

essentially equal [54]. Often a regression model is based on the number of predictors with which the AIC score is lowest. A criteria found against the AIC is that it favours sensitivity over specificity when faced with a large number of observations [54].

The Bayesian Information Criterion (BIC) was developed slightly later than the AIC by Gideon Schwarz [55]. Schwarz aimed at providing an alternative solution to the problem of choosing an appropriate model given the information received from a set of observations [55]. The one statistical difference between the BIC and Akaike’s solution is that the dimension is multiplied by ½ log n. What this means is that when faced with large numbers of observations the BIC emphasizes a parsimonious model, punishing models with more variables [55]. One advantage that BIC has over AIC is that it is consistent, meaning it will consistently select the smallest adequate model [54]. The AIC is not consistent as it may choose an overly complex model when the amount of observations is large [54]. Schwarz also denotes the possibility that Akaike’s solution cannot be asymptomatically optimal [55]. A criteria found against the BIC is that whilst penalizing models with a large number of variables, it runs the risk of under fitting the model [54].

(20)

17

Neither of these two information criteria will be perfect. AIC often risks choosing model which is too large, whilst BIC risks choosing a model too small [54]. The amount of observations is influential in the selection of information criteria, but requires information which real data will not be able to give. As cited by Dziak et al. it is suggested by Collins & Lanza, 2010, to use both information criteria to create a range of model choices, after which one is selected based on other kinds of criteria [54, 56].

2.2.7 Prediction outcome measures

One cannot assume predictions will always be correct. At the same time it is not wise to completely disregard those results that are slightly off of the true number. To aid those that try to make predictions several outcome measures have been developed.

One used most often is known as R² (R-squared). This value explains the degree to which your model explain the variation of the outcome variable [57]. For example when the R² is 0.6 the collective variables in the predictive model explain 60% of the variation in the outcome variable. The higher the

R², the better the model picked explains the variation of the outcome variable. An important

observation to keep in mind is that for every variable added to the model the R² will go up, regardless of their relationship with the outcome variable. Hence the model with the ‘best’ R² often isn’t the best predictive model able to be chosen. The adjusted R², a different outcome variable, is used most often as it adjusts for the amount of variables included in the model [58]. Meaning adding variables doesn’t necessarily result in an increased R².

Another measure used in the process of evaluating a predictive model is the Mean Squared error (MSE). This is the average of the squared distances between a fitted (regression) line and the data points in a data set. For each data point the distance is taken from the point to the corresponding fitted line. This is known as the error. The error is then squared (to avoid negative values from cancelling out positive values), added up for all data points and divided by n – 2, where n is the number of observations. If the MSE is small, it means the fitted line is close to the data and thus the predictive model handles the data well. One issue with MSE is that the unit used is the squared unit of the variable plotted on the vertical axis, which means interpretability suffers.

The Root Mean Squared Error (RMSE) solves this issue by providing a measure in the same measurement scale as the variable on the vertical axis [59]. The difference is that it reports the square root of the MSE. The RMSE can be defined as the distance between a fitted line and the data points in a data set (so not the squared distance). Literature is not entirely conclusive in how a research can best interpret the RMSE value. One suggestion is that it should not be similar to the standard deviation of the true value as this means the predicted value provides little information [60]. Higher error than a standard deviation would also be infeasible. A viable RMSE value is very much related to the context of the problem one tries to solve. Another way to ensure the RMSE value is viable is comparing two different data sets (train sample and validation sample) with each other. If the RMSE value is much higher in the validation sample it is likely that the model constructed does not provide accurate predictions. Some researchers promote the use of the Mean Absolute Error (MAE) rather than the RMSE as it provides a more natural measure of average error [61].

(21)

18

Chapter 3 Hospitalization-associated-disability: an

analysis of importance of variables in

the Hospital-ADL study

Abstract

Hospitalization-associated disability (HAD) is a condition in geriatric patients that is still largely inexplicable. Valuable insights can be found if the most influential variables of the multitude of possible underlying mechanisms are determined. Using factor analysis this study attempts to determine groups of variables which correlate highly together and have high negative correlations with all other variables, whilst explaining the most variance. Using Horn’s parallel analysis, three factors were retained. After close examination they were identified as: Physical state factor, Mental state factor and Mortality factor. Consistent with the literature each factor consisted variables clearly related to each other, representing a latent factor known for its relationship with HAD. The factors together had 20 variables which loaded significantly and explained 24.10% of variation in the dataset.

Key words: Hospitalization-associated disability, factor analysis, horn’s parallel analysis, variance,

(22)

19

3.1 Introduction

The Hospital-Associated Disability and impact on daily Life (Hospital-ADL) study is an AMC-organized multi-centre study in seven Dutch hospitals investigating HAD [5]. With the aim to find out more about HAD, the study attempts to provide answers to concerns and wishes posed in other studies. The study is unique because it is the first study where risk factors on the cognitive, behavioural, psychosocial, physical and biological levels are studied together. In addition, the study includes the use of an activity tracker to see if it provides interesting information or useful predictors for functional decline. The Hospital-ADL study consists of many different measurement moments, with testing being done at intake, during stay, at discharge and at three different follow-up moments. To further augment the already promising findings of the Hospital-ADL study, this chapter will attempt to identify:

Which variables are most important when comparing cognitive, behavioural, psychosocial, physical and biological components pertinent to functional decline together?

The chapter will also study whether common risk factors, found in other studies, are found to be important in comparison to each other. Covinsky inspected several studies examining individual risk factors for HAD [7]. The studies included connected age, depression, mobility, cognitive impairment and delirium on admission to loss of function. Comparing the combination of variables included in the Hospital-ADL may show if these individual risk factors are related and which may be more important in comparison to the others.

In addition we aim to explore the interdependencies between variables as this may show new novel relationships between established risk factors. Whilst most results are expected to be consistent with the results of other studies the fact that all different levels are studied together may provide a new frontier. We postulate that factor analysis, which is a form of statistical learning, will push the envelope for the current methodology used in functional decline. Using statistical learning allows us to analyse the large amount of variables included, takes away the potential biases in variable selection and leaves it to statistical analysis to provide answers to our questions. The questions we attempt to answer in this chapter are:

• How can a fraction of variables capture the most important information pertinent to a certain concept, whilst maintaining a comparable level of information?

• How can latent concepts introduce new novel connections between variables relevant for functional decline?

• How does the importance and severity of risk factors in functional decline relate to their loading level?

(23)

20

3.2 Methods

3.2.1 Materials

Data was obtained from the Hospital-ADL study. Each patient fitted the applied inclusion criteria for the multicentre study: 1) acute admission to the geriatric, internal medicine or cardiology department in one of the included hospitals; 2) aged 70 years or older; 3) scored a 15 or higher on the Mini-Mental State Examination (MMSE); 4) had approval from the attending medical doctor for inclusion; 5) satisfactory proficiency in the Dutch language to understand the included questionnaires. Patients were excluded when they: 1) were disabled in all six ADLs; 2) had an observed life expectancy of three months or less. Within 48 hours after admission each patient was exposed to a range of tests shown in table 1. Due to time constraints we were not able to include the activity tracker data.

Question or instrument

1. Medical and demographical data

Age Date of birth

Gender Postal code

Date and time of admission Education

Ethnicity Country of birth patient and parents

Marital status Living arrangement

Medical comorbidity CCI

Severity of acute illness MEWS

Admission diagnosis

2. Personal interviews/self-reported data

2.1 Cognitive functioning

Cognitive impairment MMSE

Delirium CAM

2.2 Behavioural and psychosocial functioning

Fear of falling NRS fear of falling

Anxiety STAI-6

Apathy GDS-15

Quality of life

1] In general, how is your quality of life?

2] How would you grade your life at this moment, with a range between 0 and 10? And;

3] Compared to one year ago, how would you rate your health in general now?

EQ-5D

2.3 ADL/Physical functioning

Disability in ADLs Modified Katz Index Scale

Independence in walking FAC

Mobility Could you walk outside for 5 minutes two weeks before admission/currently? And how often did/do you do physical activity two weeks before admission/currently?

Falls Have you fallen once or more in the past (six) month(s)? If yes, how _{many times?}

Pain NRS pain

Fatigue NRS fatigue

Sleep quality PSQI

Sleep medication PSQI

Daytime sleepiness Do you currently suffer from daytime sleepiness? If yes, does this _{affect your daily living?} Polynocturia Do you currently suffer from polynocturia? If yes, does this affect

(24)

21

your daily living?

Dizziness Do you currently suffer from dizziness? If yes, does this affect your _{daily living?} Shortness of breath Do you currently suffer from shortness of breath? If yes, does this _{affect your daily living?} Hearing impairment Do you experience difficulties with hearing, despite the use of _{hearing aid?} Vision impairment Do you experience difficulties with your vision, despite the use of _glasses?

Nutrition SNAQ

Dependency Do you smoke? Do you use alcohol?

Polypharmacy Do you use five or more different medications?

2.4 Health care utilization

Readmission Have you been hospitalized in the last (six) month(s)? If yes, for _{how many days?}

3. Physical performance tests

Handgrip strength Jamar

Mobility DEMMI

Agility CSR

Balance, strength, and gait SPPB

Walking distance 2MWT

Body composition BIA (Bodystat Quadscan 4000)

Activity tracker Fitbit Flex

4. Blood parameters Inflammation markers CRP WBC diff TNF-a IL-6 IL-8

TABLE 1: summary of outcome measures and tests of H1 in Hospital-ADL study [5]

3.2.2 Design

This study is a retrospective, statistical analysis of a cohort study designed by the geriatrics department at the Academic Medical Centre in Amsterdam (AMC). The Hospital-ADL study started in October, 2015 and is currently carrying out its last follow-up. This chapter uses the gathered data so far to make a provisional first analysis. Whilst data is available for all measurement moments, their completeness, given the many different components, is limited. Since most components are available in the first measurement moment, which is within 48 hours of admission, only that specific data is analysed.

3.2.3 Procedure and analysis

The data was provided by the team carrying out the Hospital-ADL study. The dataset included free-form questionnaire fields and variables with near-zero variance. These were consequently removed. The data consisted of quantitative data as well as qualitative data. All missing values were imputed using a principal component methodology [62]. The imputed qualitative variables, which were either dichotomous or polychotomous, were transformed to numerical values. To account for variability in unit scale the quantitative and transformed qualitative variables were appended and transformed to a correlation matrix. To ensure binary and polychotomous variables could legislate themselves a heterogeneous correlation matrix was used.

We use a common factor analysis, which is a multivariate statistical method reducing a number of variables to a smaller number of latent factors. Each factor has a specific ‘loading’ on a selection of

(25)

22

variables that together explain the latent concept it represents. Using Horn’s parallel analysis, the number of factors that were to be retained was chosen. The appropriateness of factor analysis was examined by executing the Kaiser-Meyer Olkin (KMO) test and Bartlett’s Test of Sphericity. The Kaiser-Meyer Olkin is the measure of sampling adequacy, required for confirming an adequate case to variable ratio [63]. The KMO varies between 0 and 1, with 0.60 being the accepted and recommended minimum index. The Bartlett’s test of sphericity tests if there is a redundancy between variables that can be explained by latent factors [64]. More specifically it tests the null hypothesis that the given correlation matrix is an identity matrix (a square matrix with diagonal elements equal to one and non-diagonal elements equal to zero). Promax rotation was used for enhanced interpretability of factor loadings [51].

Whilst literature is not conclusive on what a significant factor loading is, it is suggested that factor loadings between 0.40 and 0.70 are realistic [65]. Communalities of 0.80 and higher are unlikely. Lower loadings will most likely create cross-loading items, meaning they load on two or more factors, whilst requiring high loadings will result in constructs with too few items. Cross-loading items most likely show items that are poorly written or have a flawed underlying factor structure [42]. Constructs with too few items have little power and are generally weak and unstable [42]. Promoting improved interpretability of results a loading 0.50 or higher was selected, following the general rule of thumb giving by literature for having items with no cross-loadings [66, 67].

3.3 Results

Descriptive statistics of the continuous (means, standard deviations, minimums and maximums), binary (yes/no distribution) and polychotomous variables (level distribution) can be found in appendix B. The heterogeneous correlation matrix is available on request as it is too large to fit in this report.

A total of 401 patients were included in the study. Their age ranged from 70 to 99, with a mean age of 79.671 (±6.658 SD). After removal of redundant and error-prone variables each variable had, on average, 13.95% missing values. Using Horn’s parallel analysis, it was revealed that there were 3 latent factors to be retained after factor analysis. After Promax rotation, the three factors explain 24.10 % (18.07/75) of variance of the original 75 variables (table 1). The factor-variable correlations are represented in table 2, showing the contribution of each variable.

The first factor, which explains 13.95 % of the total variance, has high positive loadings on age, katz6 scores, katz15 scores and FAC score. The factor had high negative loadings on FFMI BIA, handgrip strength, SPPB score and DEMMI score. As most variables were associated with functional status this factor was thus referred to as Functional status factor. The second factor represented high loadings on GDS score, depression and the grade given to fatigue, whilst scoring high negative loadings on quality of life 2 and the EQ5D-NL test. It will be referred to as Mental state factor. The second factor accounted for 6.15% of variance. The third, and last, factor contributed for 4% of generalized variance and was associated with high loadings on the number of hours and days in the hospital, the respiratory rate and the MEWS. As most variables are connected to mortality risk factors the factor will specified as the Mortality factor.

(26)

23

The KMO statistic of 0.74 exceeded the required adequacy index (0.60) and Bartlett’s test of sphericity (Chi-square = 19652.1; p < 0.01) suggests that the factor analysis was applicable to the data set and latent factors were present. The sample size requirement, or case to variable ratio before factor analysis (5.35 to 1), passed the minimum 5 to 1 standard ratio.

Variables

Factor 1 Factor 2 Factor 3

Functional status (physical factors) Mental state (psychosocial factors) Mortality Age -0,53643527 -0,159260386 0,041853358

Number of hours spent in

hospital 0,036941104 0,102891427 0,685427351 Number of days admitted pre 0,036978923 0,102860165 0,685441412

Respiratory rate 0,125476597 0,123336536 0,599108401

MEWS score 0,074814251 -0,119821192 0,633261211

Katz6 score premorbid -0,581920559 0,044276575 0,224899641 Katz6 score -0,631890927 0,09726497 0,324314909 katz15 score premorbid -0,753627999 0,06263371 0,246334787 katz15 score -0,744463707 0,111280957 0,32218771 Fatigue grade -0,057861721 0,5464509 -0,084508462 GDS score -0,06249732 0,794958225 -0,165775313 Quality of life 2 -0,029602187 -0,611665736 0,051662559 EQ5D-NL 0,26399057 -0,544032289 -0,112008642 FFMI BIA 0,561146409 0,057957743 0,331344963 Handgrip strength right 0,737631247 0,01270715 0,110845008 Handgrip strength left 0,763781469 0,021273749 0,078403677 SPPB score 0,613230658 -0,119293538 -0,340659787 DEMMI score 0,563441372 -0,123616908 -0,396511669 Depression -0,012076798 0,691614973 -0,117594374 Respiratory rate 0,025868157 0,182834428 0,579220636

Table 2: factor coefficients

3.4 Discussion

The aim of this study was to examine which variables, included in the Hospital-ADL study, are most important. Factor analysis caused the number of variables to be reduced by 70.66%, and managed to identify concepts that were, consistent with the literature, closely related to functional decline. We were not able to identify novel connections between variables significantly loading in the factor analysis.

It is highly likely that factor analysis has been used before with variables pertinent to HAD, yet this analysis is in many ways novel. One key difference with most related studies is that a combination of components concerning HAD were included. The latent concepts identified by the factor analysis could therefore show unique connections to be considered between the different components relevant for HAD. Whilst the latent factors did not reveal any new unlikely connections, it did present variables connecting to each other. Consistent with literature the performance based measures of functional status (DEMMI, SPPB, handgrip strength) were, even though measured in a different way, related to self-reported measures of functional status (Katz) in the first factor [68]. Age, which is a heavy influencer of HAD, was also included with the physical state factor [7]. We can explain this by its direct relation to the physical function, regardless of the influence of other components found in the data. The psychosocial variables combined in the second factor are likely linked together in literature. Whilst outside the scope of the study the relationship between quality of life and

(27)

24

depression is apparent. Consistent with these observations the EQ5D-NL, which measures the health-related quality of life, is also included [69]. The last, and most difficult to interpret, factor includes seemingly unrelated variables. Yet when observing the Modified Early Warning Score (MEWS) is included, which serves as a measure of severity of illness in a patient, it becomes more apparent. Each separate variable is in some way related to a generalised likelihood of death. Respiratory rate and time spent in the hospital are both indicators of a heightened chance of dying [30, 33].

3.4.1 Strengths and weaknesses

The key strength of this study as compared to other studies is the novel possibility of analysing all relevant components together. In contrast other studies mostly research and validate a small number of risk factors at a time or compare a multitude of studies in an attempt to combine them through meta-analysis or literature research [7, 11, 14, 17, 28, 29]. In addition the amount of recruited acutely hospitalized elderly is substantial given the considerable challenges facing researchers wanting to study the acutely ill elderly [70]. The number of variables included in the data set also introduced potential weaknesses for analysis as it included a large percentage of missing values. In addition, it contained several different data types to be analysed together (dichotomous, polychotomous and continuous), which further complicated the statistical analysis by necessitating a heterogeneous correlation matrix. The researchers of the Hospital-ADL study have had to manually input many of the variables, which introduces a substantial chance of human error in the contents of the variables. The potential errors caused by these three limitations made the analysis more complicated but did not have a significant lasting effect on the results.

3.4.2 Clinical relevance

The study demonstrates its clinical relevance in the identification of interdependencies between given variables pertinent to functional decline. These interdependencies diminish the amount of relevant variables to a fraction of what was initially used whilst preserving a viable and comparable level of information applicable to functional decline. In research this positively affects the viability for prediction in a clinical setting as less variables are required. A workable outcome measure should still be picked before a prediction model can be constructed. The study primarily showed the statistical feasibility for the factor analysis methodology for reducing the dimensionality of the data, but it could benefit from an experts opinion by a medical doctor to draw conclusions relating to the content.

3.4.3 Future work

The fact that the results were consistent with primary elements of functional decline is sensible as the variables in the Hospital-ADL study were picked to capture the most information pertaining to functional decline. Physical and psychosocial variables, specifically those related to functional decline, are found together in the first two factors. The biological, cognitive or behavioural components are much less prominently included. Only a few exceptions such as age and respiratory rate remained. This observation could mean these components are much less important given the context of functional decline. In addition the major risk factors cognitive impairment and delirium on admission were not captured in the identified latent factors. Whilst variables representing these risk factors were included in the analysis they did not load significantly enough in comparison to other

(28)

25

variables. This could mean their severity or significance to functional decline could be less in comparison to age, mobility and depression. Even though it was not the primary objective the study showed a glimpse of how each variable relates to each other and compares in importance concerning the overarching theme of functional loss. A closer look at whether physical and psychosocial components to HAD are more important than others and the missing major risk factors in cognitive impairment and delirium on admission is warranted.

3.5 Conclusion

Key points

• Using factor analysis, the dimensions of the dataset can successfully be lowered to a few interpretable latent variables.

• Physical and psychosocial components of functional decline may be considered the primary indicators for functional loss.

• The factor analysis loading levels of each variable can be seen as a performance or severity indicator for that specific variable given the context of functional decline.

• The identification of interdependencies between variables show promise in showing new connections between variables relevant for functional decline.

• Using factor analysis positively affects the viability for using a prediction model in the context of functional decline.

Using factor analysis has proven to be a viable statistical method to reduce the number of variables pertinent to functional decline. The latent factors accurately portray most of the major risk factors as given by the literature. The factor analysis allows for a multitude of other observations to be made concerning the primary components of functional decline, the severity of each risk factor in comparison to each other and the identification of novel connections between variables relevant for functional decline. Further improvement of clinical care could primarily lie in the effective application of predictive modelling using factor analysis to diminish the dimensionality of the data.

(29)

26

Chapter 4 Hospitalization-associated-disability:

predicting functional loss using Katz

index of independence in activities of

daily living

Abstract

Knowing the mechanisms behind what predicts the risk of losing an activity of daily living (ADL) during a patients stay at the hospital is important for the development of viable care processes. The determination of relationships between the Katz score, which is the index of independence in ADL’s, and characteristics related to functional loss may provide necessary information for internist working with geriatric patients. Using a combination of factor and multiple regression analysis this study attempts to determine the relationships between Katz score and 71 possible functional status characteristics. First, factor analysis was utilized to reduce the number of possible predictor variables and simplify the relationships among the given variables. Then, using Horn’s parallel analysis, three factors were selected and implemented as independent predictor variables in multiple regression. It was found that a combination of the three independent factors plus six other variables had significant effects on Katz score. Together the factors explained 22.31% of variation in Katz score.

Key words: Katz index of independence in activities of daily living, factor analysis, multiple

(30)

27

4.1 Introduction

To determine when hospitalization-associated disability (HAD) will most likely occur in an elderly patient is both valuable and difficult. Especially as ADL functioning in elderly is a highly predictive indicator of mortality and hospital outcome [14, 16, 30, 33]. The Hospital-ADL study provides extensive data, which allows analysis, such as prediction, to take place [5]. In this chapter we attempt to identify the, potentially novel, risk factors that can be used to predict functional decline in elderly. The variables that have already proven itself as predictors are especially of interest in order to see if they remain consistent with their expectations. Examples of such variables, that are included in the Hospital-ADL study, are age, depressive symptoms, dementia, albumin levels, mobility, socio-economic status and cognitive impairment [14, 17, 18, 24, 25, 28, 30, 71].

The Katz score, which is an altered version of the Index of Activities of Daily Living, is used as the outcome variable [6]. Nielsen et al. suggested that there is significant difference between self-reported and performance based functional status tests [68]. Both kinds of tests are included in the Hospital-ADL study. Due to time-constraints we were not able to use both in a regression analysis. Therefore only the Katz score is used as an outcome variable. The question we attempt to answer in this chapter is:

How can the Katz score (functional loss) be predicted using cognitive, behavioural, psychosocial, physical and biological components?

McCusker et al. suggests that meta-analysis and comparison is difficult due to the considerable methodological variability between studies [39]. We postulate that the available variables, provided by the Hospital-ADL study can be used to create a clinically usable predictive model for outcome measures relevant for functional loss (Katz score), whilst avoiding the need for meta-analysis between studies.

4.2 Methods

4.2.1 Data collection

Data was obtained from the Hospital-ADL study. Each patient fitted the applied inclusion criteria for the multicentre study: 1) acute admission to the geriatric, internal medicine or cardiology department in one of the included hospitals; 2) aged 70 years or older; 3) scored a 15 or higher on the Mini-Mental State Examination (MMSE); 4) had approval from the attending medical doctor for inclusion; 5) satisfactory proficiency in the Dutch language to understand the included questionnaires. Patients were excluded when they: 1) were disabled in all six ADLs; 2) had an observed life expectancy of three months or less. Within 48 hours after admission each patient was exposed to a range of tests shown in table 1.

4.2.2 Design

This chapter is a retrospective, statistical analysis of a cohort study designed by the geriatrics department at the Academic Medical Centre in Amsterdam (AMC). The Hospital-ADL study started in October, 2015 and is currently carrying out its last follow-up. This study uses the provided data to

(31)

28

create a predictive model for functional decline. For this it exclusively uses the first measurement moment, which was within 48 hours of admission.

4.2.3 Variables of interest

The outcome of interest was the Katz, defined as the index of Independence in Activities of Daily Living. The Katz-ADL, as it is commonly referred to, is one of the more appropriate methods to determine function status using the patients ability to perform activities of daily living (ADLs) independently [6]. It is most effective when used on older adults in a variety of care settings. Several categories of explanatory variables were considered. On a patient-level medical and demographical data were examined, including age and severity of illness. The self-reported variables included cognitive functioning, behavioural and psychosocial functioning and physical functioning. Each with a variety of variables. Next, health care utilization, physical performance tests and blood parameters were considered. All covariates, including those initially coded as categorical variables, were transformed and coded as numerical variables. Missing values were imputed.

4.2.4 Procedure and analysis

The main approach is to use multivariate statistics in which factor analysis is used to classify predictor variables according to their interdependencies and to predict functional loss in elderly patients. In accordance with what McCusker et al. suggested primary components of HAD were combined through factor analysis [39]. In addition the factor analysis reduced the dimensions of the dataset. Whilst it was suggested that prehospital functioning could be a predictor for current functioning, we decided to not include these scores as they were combined into a single outcome measure and were too closely related to the Katz scores used as the outcome variable [38]. Using the Katz score as the outcome variable meant they had to be removed from the regression model.

Both backwards and forwards selection were employed to determine which model provides the best fit, as the amount of variables made best fit selection infeasible. For each number of variables the possible models were analysed concerning their BIC, Cp (AIC equivalent) and adjusted R-squared statistics. The BIC is taken as the primary indicator seeing as the number of observations and variables is high [54]. Using the provided statistics the best model was selected based on its lowest BIC value and the adjusted R-squared statistic with the best interpretation/simplicity trade-off. The Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) were given for the provided predictive model to determine its accuracy. This is consistent with the suggestion by Collins & Lanza [56].

4.3 Results

Descriptive statistics of the continuous (means, standard deviations, minimums and maximums), binary (yes/no distribution) and polychotomous variables (level distribution) can be found in appendix B. Horn’s parallel analysis determined 3 latent factors. After Promax rotation, the three factors explain 22.31 % (15.84/71) of variance of the original 71 variables. The factor coefficients of the rotated factors show the relative contribution of each trait to a particular factor (table 2).

(32)

29

Variables Factor 1 Factor 2 Factor 3

Mental state Body composition Mortality

Number of hours spent in hospital 0,267457 0,178466 0,639785

Number of days admitted pre 0,267414 0,178499 0,639792

Respiratory rate 0,179118 0,168085 0,640017 MEWS Score -0,06205 0,094526 0,710609 GDS Score 0,768973 -0,05859 -0,07611 Apathy score 0,518906 -0,13679 -0,04229 Quality of life 2 -0,578 -0,03611 -0,00071 EQ5DNL -0,67882 0,164055 -0,17603 BIA Length -0,01236 0,547722 0,061113 BIA Weight 0,046867 0,54318 0,085895 FFMI BIA -0,00949 0,687098 0,095148

Handgrip Strength Right -0,23821 0,713231 -0,07145

Handgrip Strength Left -0,23542 0,746247 -0,12279

Depression 0,670699 0,014314 -0,07199

Respiratory rate 0,273353 0,078802 0,647417

Table 1: rotated factor loadings

The first factor, explaining 11.82% of the generalized variance, was characterized by high positive loadings (factor-variate correlations) on the geriatric depression scale (GDS), apathy score and depression. The factor has high negative loadings on quality of life 2 and the EQ5D-NL test. As most factors are associated with psycho-social and mental components it is termed Mental state factor. The variables, all positively loaded, associated with the second factor were BIA length and weight, FFMI BIA and handgrip strength left and right. Contributing to 6.45% of variance, the factor will be referred to as Body composition factor. The third and last factor, responsible for 4.05% of variance, had high loadings on the number of hours and days stayed in the hospital in the three months prior to admission, respiratory rate and the MEWS. As most variables are connected to mortality risk factors the factor will specified as the Mortality factor. A total of 15 variables were reduced to three factors explaining Mental state, Body composition and Mortality accordingly.

The KMO statistic of 0.71 exceeded the required adequacy index (0.60) and Bartlett’s test of sphericity (Chi-square = 17232.41; p < 0.01) suggests that the factor analysis was applicable to the data set and latent factors were present. The sample size requirement, or case to variable ratio before factor analysis (5.65 to 1), passed the minimum 5 to 1 standard ratio.

Factor coefficients were used to obtain factor score values. These values for the three selected factors were used as independent variables in multiple linear regression analysis to determine significant factors and variables for Katz score. Using visual aids and backwards selection (figure 1) range of models was constructed and compared in accordance to literature [56]. Backwards selection was chosen as forwards selection gave an equivalent result and best selection, with this much variables, was infeasible. Adjusted R-squared and Mallows’ Cp (equivalent to AIC) both gain little after the ninth variable Primarily leaning on the BIC it was determined that 9 variables would provide the best model.

(33)

30

Figure 1: visual aids (top left: Adjusted R², top right: Mallows’ Cp and bottom: BIC)

A multiple linear regression was calculated to predict the Katz score based on the variables presented in table 2. A regression equation was estimated (F(9,391) = 88.37, p < .000), with an R² of .6704. Participants predicted functional loss is equal to -0,27017 (INTERCEPT) + 0,003354 (SYSTOLIC BLOOD PRESSURE) - 0,05896 (SPPB SCORE) + 0,198483 (HOSPITALIZATION) - 0,29085 (FATIGUE) + 0,136912 (LIVING SITUATION) + 0,102815 (MOBILITY SCORE 1) + 0,339647 (MENTAL STATE) - 0,34635 (BODY COMPOSITION) + 0,183714 (MORTALITY). Systolic blood pressure ranges from 69 to 220, SPPB score ranges from 0 to 12, Hospitalization is coded as 1 = No, 2 = Yes, Fatigue is coded as 1 = No, 2 = Yes, Living situation is coded as 1 = independent alone/with others, 2 = service flat, 3 = care home, 4 = nursing home/revalidation centre, and Mobility score 1 ranges from 1 = no difficulty to 5 = not possible. Mental state, body composition and mortality are all composite factor scores for a multitude of variables found in table 1. The functional loss of patients increased with a higher systolic blood pressure, a lower fatigue, a more dependent living situation, a worse mobility score, a worse mental state, a worse body composition and a higher mortality score. All given variables were significant predictors of functional status. All of the selected factors were found to have a significant linear relationship with the Katz score (p<0.001).

(34)

31

Coefficients regression Estimate Std. Error t value Pr(>|t|)

(Intercept) -0,27017 0,252537 -1,06983 0,285354

Exact systolic blood pressure 0,003354 0,00126 2,662421 0,008078

Total score SPPB -0,05896 0,012759 -4,62104 5,19E-06

Hospitalization 0,198483 0,060831 3,262855 0,0012

Fatigue -0,29085 0,074899 -3,88325 0,000121

Living situation 0,136912 0,042577 3,215625 0,00141

Mobility score 1 0,102815 0,024716 4,159764 3,92E-05

Mental state 0,339647 0,042973 7,903768 2,78E-14

Body composition -0,34635 0,035515 -9,75206 2,96E-20

Mortality 0,183714 0,034531 5,320267 1,75E-07

Table 2: Multiple regression coefficients

Using the regression model predictions for the Katz index were made. Descriptive statistics for the Katz score can be found in table 3. The mean absolute error (MAE) for the model was 0.438922, with a root mean squared error of 0.5554245. The R² was 0.670414. The values for BIC and the AIC were 732.3281 and 688.39945 respectively. An overview of all outcome statistics can be found in table 4.

Variable Mean SD Min Max

Katz score 1.163368e-16 _0.9686856 _-1.254013 _3.011109

Table 3: Descriptive statistics Katz score

Statistic Best selection model Root Mean Squared Error (RMSE) 0.5554245 Mean Absolute Error (MAE) 0.4389222

R² 0.670414

Number of variables 9

BIC 732.3281

AIC 688.3945

Table 4: Outcome statistics of best selection model

4.4 Discussion

The aim of this study was to find out how the Katz score could best be predicted using all available components given in the data set. Using a combination of regression analysis and factor analysis, a fraction of the available variables is able to predict the Katz score within, on average, half a standard deviation of the actual value.

The factor analysis successfully reduced the dimensions of the dataset by combining several interdependent variables to a three independent variables usable in multiple regression. This in combination with the unique characteristic of the dataset made this research novel in comparison to other research in hospitalization-associated disability. We provided a small selection of variables from a list of multiple proven risk factors which viably predicted the Katz score, showing not all risk factors are viable predictors of the condition.

(35)

32

4.4.1 Strengths and weaknesses

A major opportunity was found when many proven theoretical predictors were combined in the Hospital-ADL study [5]. When creating a predictive model it selects the most important variables from the group that best represent the whole resulting in a selection of the most important predictors, from a list of most theoretical predictors, representing the leading strength of the study. This is much more viable in comparison to other studies where risk factors are sparsely analysed together [7, 11, 14, 17, 28, 29].

There is no agreed upon gold standard in the assessment of functional decline [68]. Tests are either self-reported, in which the patient reports their functional capabilities in a simple questionnaire, or performance-based, during which the patient is asked to do a specific physical test pertinent to functional performance. The self-reported tests available are used most frequently due to their low time-consumption in comparison to performance-based assessments. Unfortunately only using self-reported testing introduces discrepancies and bias [68]. The outcome of interest in this study, the Katz score, is a self-reported measurement. Due to time constraints we were not able to create or find a different viable outcome measurement at admission, which introduces weakness. In addition we had to combine the 4, Katz related, variables together. This took away the advantage of the correlation matrix in that it represented the unit scale of the variable, and thus interpretability suffered. Another weakness of the data was that it was not able to provide all relevant variables which were given in the first measurement moment at any of the follow-up moments. This meant we were unable to construct the model using the admission data and unable to predict Katz at discharge or any of the follow-ups. This may have been more valuable information to the geriatric department of the AMC

4.4.2 Clinical relevance

The study demonstrates a clear clinical relevance in the ability to predict a patient’s functional decline. This knowledge can determine the further trajectory necessary to prevent further decline and adverse outcomes [31]. The model provides a first step in the hospitals ability to identify high-risk patients, which is an important target for the prevention of functional decline [11]. Steps to take after risk stratification or to alleviate further functional decline are still being researched and developed [25, 26, 72, 73]. Using studies such as the Hospital-ADL study to improve care for high-risk patients is important as a next step [68].

4.4.3 Future work

A composite physical score, similar to what is found in the factor analysis and containing all interdependent variables pertaining to the physical attributes of a patient, may provide a better indication of the full spectrum of consequences that low functional status have. In addition it negates the negative effects of using a self-reported variable such as the Katz score as it combines it with performance based physical tests. The variable itself would become of complex nature and would thus require a closer look for it to be valid in clinical use.

Determining whether an elderly patient will incur function decline during his or her complete stay at the hospital, rather than only at admission, may be more viable prediction. It means the

Hospitalization-associated disability: using statistical learning to reduce the dimensionality of data and promote analysis