Building a classifier for detecting patients with low decreasing QoL trajectory The aim is to build a binary classifier able to detect patients at risk of low QoL trajectory during the 18 month period from baseline. The positive class is the low decreasing QoL trajectory group identified in the first part of the analysis (Fig. I12). The negative class emerges from the grouping of the high increasing and moderate QoL trajectory, as depicted in Fig. I12. It is noted that the specific group of patients are the least resilient group based on both QoL and depression trajectories. The vast majority of these patients also belong to the “Poor” depression trajectory class and their mean depression levels are relatively high throughout the 18 months after diagnosis (Fig. I4). The classifier is trained on 513 patients. Predictors considered are psychometric scales, sociodemographic, clinical, medical and treatment variables from baseline or baseline and M3. M6 is also considered to evaluate performance improvement. The number of predictors is 96 from M0 and 169 from M0 and M3. A random forest classifier was chosen for building the final models. Figure I12: The considered classes The distribution of the trajectory classes between the clinical sites is reported in Table I3.
Building a classifier for detecting patients with high improving QoL trajectory The aim is to build a binary classifier able to detect patients of very high QoL trajectory during the 18 month period from baseline. The positive class is the high increasing QoL trajectory group identified in the first part of the analysis (Fig. I12). The negative class emerges from the grouping of the low decreasing and moderate QoL trajectory, as depicted in Fig. I15. It is noted that the specific group of patients are the most resilient group based on both QoL and depression trajectories. The vast majority of these patients also belong to the “good” depression trajectory class and their mean depression levels are very low throughout the 18 months after diagnosis (Fig. I4). The classifier is trained on 513 patients. Predictors considered are psychometric scales, sociodemographic, clinical, medical and treatment variables from baseline or baseline and M3. The number of predictors is 96 from M0 and 169 from M0 and M3. A random forest classifier was chosen for building the final models.
