T51 - Using Machine Learning To Build Individualized Prediction Models Of Future Quality Of Life In Psychosis Patients

Abstract

Background Schizophrenia (SCZ), bipolar disorder (BD) and schizoaffective disorder (SCZA) are severe, chronic mental illnesses, characterized by a marked decrease in quality of life (QoL) and functioning in a great proportion of patients. QoL, subjective well-being and psychosocial functioning are important determinants of patient satisfaction and patient-physician relationship and improvement in these domains is often more important for the patients than the reduction of clinical symptoms. With the growing emphasis on personalized care in everyday clinical practice, there is an increasing need for tools that can provide individual predictions on the future course of these patient-centered domains and thus help clinicians in providing individually tailored interventions. Therefore, we used machine learning to predict general and psychological QoL 6-months after an initial assessment in individuals diagnosed with SCZ, BD and SCZA using a combination of clinical, demographic and genetic variables. Methods In an ongoing longitudinal naturalistic study (www.kfo241.de, www.PsyCourse.de), patients meeting DSM-IV criteria for BD, SCZ and SCZA were recruited at multiple sites across Germany and Austria. Information on their sociodemographic background, family history, current psychiatric symptoms, functioning and QoL (WHOQoL-BREF) was obtained with a battery of rating scales in 6-month intervals. For the current analysis results of T1 (N=764) and T2 (N=428) were used. Participants were genotyped on the PsychChip (Illumina) whole-genome SNP array and their SCZ polygenic risk scores (PRS) at 11 different p-value thresholds (pTs) (pT1=0.00000005, pT11=1) were calculated using data from SCZ PGC2 (excluding German participants) as training data set. The importance of the clinical and demographic features and the cumulative SCZ polygenic risk in predicting 6-month (T2) general and psychological QoL was tested using linear Support Vector Machines. The prediction algorithm was wrapped into a repeated-nested cross-validation setting to ensure good generalizability. Results Higher baseline scores on the negative symptoms subscale of the PANSS, higher levels of baseline depression (BDI), lower level of functioning (GAF) and being unemployed or having a part-time job were the most important determinants of impaired T2 general or psychological QoL (test-fold balanced accuracy 61.2% and 71.3%, respectively). The PANSS negative symptoms subscale items were more important in the prediction of general QoL, whereas items of the BDI depression scale had higher weight in predicting psychological QoL. Though having higher SCZ-PRS were associated with lower T2 general QoL, compared to the clinical variables, the SCZ polygenic risk seemed to play a less important role in determining future QoL and thus was not included in the final prediction model. Discussion Our results indicate that prognostic tools using clinical data can potentially be used to indicate future general and psychological QoL of SCZ, BD and SCZA patients. However, the predictive accuracy requires improvement to be clinically useful. Further research will investigate the association found here between SCZ-PRS and lower general QoL with more sensitive biological markers. Specifically, we are currently working on techniques which would allow us to add further genomic or transcriptomic data to the analysis.