Use case analysis

The following section is based on the dimensionality reduction output that the neonatal experts found interesting, although they also looked at the cluster summary graph and the patient detail, there was not enough time to get specific conclusions from the user-test. It also appeared that the test users got stuck after exploring the cluster overview component. We will show one case for maturation analysis and one case for crash moment analysis. The dimensionality reduction outputs are recreated with the same inputs as used by the neonatal experts. We use UMAP with the hyperparameters on n-neighbors: 3 and min-dist: 0.04 as optimal settings.

Maturation analysis

For the maturation analysis we copy the settings that one neonatal expert used:

• parameters HF and SpO2

• GA range of [24, 32] weeks

• Time series range in days [1, 14]

We use UMAP to retrieve the images in figure 7.12and figure7.13, with distances Dynamic Time Warping and RMSE respectively. As most of our outputs in maturation analyses, the pro-jection colored by gestational age show that the NICUdash finds a separation between gestational age, although it does not display this with a high contrast. Figure7.12and7.13both show scatter points of sepsis patients together clearly, indicated by the blue circle in the figures. Upon selecting these scatter points, it appears that for both using distance measure Dynamic Time Warping and RMSE finds the same patients. Upon further exploring of this group we can see some interesting results, see figure7.14. For heart frequency (HF) the time series seem more or less similar in the days [8 - 13] which they intersect in time points. Besides, the graph for SpO2 indicates that this group is highly similar in SpO2 measurements. Now we go back to the first observation, that RMSE and Dynamic Time Warping strongly suggest grouping of these instances. It confirm that with averaging based on days, which we covered in chapter6.4, can make RMSE perform similarly to Dynamic Time Warping.

Figure 7.12: Maturation analysis plot on the left colored by gestational age and on the right colored by group with a distance measure of Dynamic Time Warping

CHAPTER 7. RESULTS AND EVALUATION

Figure 7.13: Maturation analysis plot on the left colored by gestational age and on the right colored by group with distance measure of RMSE

Figure 7.14: Selected group of patients from figure7.12and7.13

Anchor point analysis

This section will cover the use case when the user wants to explore the patterns when the time series are aligned by the crash moment. While doing the crash moment analysis the neonatal expert selected the following settings:

• parameters HF and RF

• GA range of [24, 32] weeks

• Time series range in hours [-24, 24], aligned on crash moment.

Different from the previous use case, is that we will use a different distance measure in this use case. Additionally, we are aligning the time series on the known crash moment. We use a the Max-dist distance measure with UMAP. Figure7.15is the result of this input, and does not show an interesting plot for the test user. Then the neonatal expert reduced the GA range to [28, 32].

CHAPTER 7. RESULTS AND EVALUATION

This gave an output with some clear separation of control group, see figure 7.16, which with a group of scatter points on the top right and in the center bottom as the control group patients.

Figure 7.15: Dimensionality result of parameter HF and RF, GA range of [24, 32] and time series range in hours [-24, 24]

Figure 7.16: Dimensionality result of parameter HF and RF, GA range of [28, 32] and time series range in hours [-24, 24], on the top right appears a separated group of control patients, while on the bottom center another separated group of controls is located

When we try to add the SpO2 parameter, as in figure 7.17the separation disappears and it appears that SpO2 adds noise in this case.

From here on wards, we use our own interpretation as we did not have enough time with the neonatal experts to come to conclusions. We will reason with the distance measure used and return to the input without SpO2. We select the two groups of control patients, as indicated by figure7.18, and see that dimensionality reduction with the parameters HF and RF that there is a difference between the two groups. This is clearly seen at the RF graph; we also see that the spread of both these selected clusters are consistently small. This could be the result of using the Max-dist distance, which in our case returns the maximum distance it can find at a certain time point between two time series. So in this case NICUdash can distinguish between two groups of patients that are relatively stable over time and are significantly different from each other in measurement values.

We also tested with the RMSE on the settings HF, RF with a GA range of [28, 32], and we could find the same two groups of patients easily. Figure7.19, shows the result and selected groups of patients. The curves of the two groups are similar in shape, however it seems that the groups are closer and more tight when using RMSE in this case.

On the contrary, using the distance measures Dynamic Time Warping does not show the same group control group separation. For Dynamic Time Warping a possible explanation could be that the time series are already aligned, so using Dynamic Time Warping could be less effective.

CHAPTER 7. RESULTS AND EVALUATION

Figure 7.17: Dimensionality result of parameter HF, SpO2 and RF, GA range of [28, 32] and time series range in hours [-24, 24] there appear no distinguishable groups of similar patients

We can observe that when the time series are aligned by an event, the RMSE and Max-dist distances can find similar groups of instances.

Figure 7.18: View of NICUdash with settings that one neonatal expert used, selected are two groups of control patients. Using UMAP with Max-distance

CHAPTER 7. RESULTS AND EVALUATION

Figure 7.19: View of NICUdash with settings that one neonatal expert used, selected are two groups of control patients. Using UMAP with RMSE