Bagherzadeh-Khiabani, F.

Prediction of health-related quality of life by machine learning from longitudinal self-reported symptom patterns in adult survivors of childhood cancer
Farideh Bagherzadeh-Khiabani, Kevin Krull, Greg Armstrong, Melissa Hudson, Leslie Robinson, Yutaka Yasui, I-Chan Huang

Survivors of childhood cancer are at risk of various late effects. Predicting health-related quality of life (HRQoL) from longitudinal patterns of patient-reported symptoms allows early detection of high-risk subgroups and offers opportunities for interventions tailored to specific symptom patterns.

Study population involved 732 adult 5-year survivors of childhood cancer participating both Childhood Cancer Survivor Study and St. Jude Lifetime Cohort Study. Outcomes included 10 HRQoL measures including physical and mental component summary scores (PCS and MCS) and eight specific HRQoL indicators involved in their calculation, from the survey administered around 2015. Potential predictors were demographic, cancer diagnosis/treatment, and 37 symptoms, categorized into 10 domains, collected across three time-points, around 2000, 2009 and 2013. We generated longitudinal symptom-patterns hypothesized to affect HRQoL including development/disappearance or count increase/decrease between two consecutive assessment-points, and consistent presence/absence over two/three time-points (528 patterns). We used elastic-net for regularized linear regression and Bayesian Information Criterion (BIC) to select its best tuning parameters. We employed Intra-Class Correlation (ICC) and Area Under the Curve (AUC) for dichotomous HRQoL status (impaired vs. non-impaired) as performance measures, estimated by 10-fold cross-validation (CV).

For MCS, by including symptom-patterns, the CV-based ICC improved from 0.05 to 0.49 and the CV-based AUC from 0.60 to 0.82. For PCS, these improved from 0.10 to 0.48 and from 0.61 to 0.82, respectively. Results from sensitivity analysis predicting eight specific HRQoL indicators confirmed the similar substantial improvements. Also, consistently suffering/being free of symptoms or symptom of a domain appears highly predictive of HRQoL. Furthermore, selected predictors did not include soci-demographic or cancer diagnosis/treatment variables, emphasizing the importance of symptom-patterns.

Our analysis shows that longitudinal symptom-patterns successfully predict HRQOL scores in childhood cancer survivors. The availability of patient-reported data, self-perceived quality of life scores and machine learning methods paves the way towards personalized medicine approach towards QoL.