QOL-22. MACHINE-LEARNING INFERENCE MAY PREDICT QUALITY OF LIFE SUBGROUPS OF ADAMANTINOMATOUS CRANIOPHARYNGIOMA. (4th December 2020)

• Record Type: Journal Article
• Authors: Hengartner, Astrid C; Prince, Eric; Staulcup, Susan; Vijmasi, Trinka; Souweidane, Mark; Jackson, Eric M; Johnston, James M; Anderson, Richard C E; Naftel, Robert P; Grant, Gerald; Niazi, Toba N; Dudley, Roy; Limbrick, David D; Ginn, Kevin; Smith, Amy; Kilburn, Lindsay; Jallo, George; Wilkening, Greta; Hankinson, Todd
• Is Part Of: Neuro-oncology. Volume 22(2020)Supplement 3
• Page Start: iii435
• Page End: iii435
• Publication Date: 2020-12-04
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noaa222.684
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: BACKGROUND: Due to disease and/or treatment-related injury, such as hypothalamic, visual, and endocrine damage, quality of life (QoL) scores after childhood-onset Adamantinomatous Craniopharyngioma (ACP) are among the lowest of all pediatric brain tumors. Decision-making regarding management would be aided by more complete understanding of a patients likely QoL trajectory following intervention. METHODS: We retrospectively analyzed caregiver and patient-reported QoL-instruments from the first 50 patients (ages 1–17 years at diagnosis) enrolled in the international Advancing Treatment for Pediatric Craniopharyngioma (ATPC) consortium. Surveys included 205 pediatric-relevant questions and were completed at diagnosis, and 1- and 12-months following diagnosis. Using Multiple Correspondence Analysis (MCA), these categorical QoL surveys were interrogated to identify time-dependent patient subgroups. Additionally, custom deep learning classifiers were developed using Google's TensorFlow framework. RESULTS: By representing QoL data in the reduced dimensionality of MCA-space, we identified QoL subgroups that either improved or declined over time. We assessed differential trends in QoL responses to identify variables that were subgroup specific (Kolmogorov-Smirnov p-value < 0.1; n=20). Additionally, our optimized deep learning classifier achieved a mean 5-fold cross-validation area under precision-recall curve score > 0.99 when classifying QoL subgroups at 12 month follow-up, using only baseline data. CONCLUSIONs: This work demonstrates the existence of time-dependent QoL-based ACP subgroups that can be inferred at time-of-diagnosis via machine learning analyses of baseline survey responses. The ability to predict an ACP patient's QoL trajectory affords caregivers valuable information that can be leveraged to maximize that patient's psychosocial state and therefore improve overall therapy.