Joeky Senders

90 Chapter 5 Outcome and input features Although machine learning provides a variety of predictive algorithms, most of them are developed to accommodate binary or continuous outcomes instead of censored survival outcomes (i.e., time-to-event data). To facilitate a vis-à-vis comparison between traditional statistical and novel machine learning algorithms, we compared all algorithms in their ability to predict one or more of the following survival outcomes: (i) continuous: overall survival from diagnosis to death in months, (ii) binary: one-year survival probability, and (iii) censored: subject-level Kaplan-Meier survival curves. All demographic, socio-economic, radiographic, and therapeutic characteristics available at individual patient-level in the SEER registry were included as input features. Continuous variables included age at diagnosis (years) and maximal enhancing tumor diameter in any dimension (millimeters). Categorical variables included sex, race (White, Black, Asian, other), ethnicity (Hispanic, non-Hispanic), marital status (married, non- married), insurance status (insured, uninsured/Medicaid), tumor laterality (left, right, midline), tumor location (frontal, temporal, parietal, occipital, cerebellum, brainstem, ventricles, overlapping lesion), tumor extension (confined to primary location, ventricle involvement, midline crossing), surgery type (biopsy, sub-total resection, gross-total resection), and administration of any form of postoperative chemotherapy and/ or radiotherapy. Data on input features and survival outcomes were collected by independent, trained data collectors. Statistical analysis Missing data was multiple imputed by means of a random forest algorithm. 5 The total cohort was randomly split into a training and hold-out test set based on an 80/20 ratio. The Cox proportional hazards regression (CPHR) and the Accelerated Failure Time (AFT) algorithms allow for inferential analysis on censored survival data. Therefore, both approaches were also utilized to provide insight into the independent association between covariates and survival. Interactions between age, sex, surgery type, radiotherapy, and chemotherapy were modeled in both approaches. The Benjamini-Hochberg procedure based on 41 comparisons (26 parameters plus 15 two- way interactions) was used to adjust for multiple testing. The proportional hazards assumption of the CPHR model was assessed by means of the Schoenfeld Residuals Test, and the distribution assumption of the AFT by means of a quantile-quantile plot. All covariates that were statistically significantly associated with survival in the inferential analysis were included in the predictive analysis. For the predictive analysis, 15 machine learning and statistical algorithms were trained including AFT, bagged decision trees, boosted decision trees, boosted