Distillation of the clinical algorithm improves prognosis by multi-task deep learning in high-risk Neuroblastoma

• DOI: 10.1371/journal.pone.0208924
• Publication Date: 2018-12-07T18:25:30Z
• Authors: Valerio Maggio; Marco Chierici; Giuseppe Jurman; Cesare Furlanello
• Keywords: Male; 0303 health sciences; Science; Gene Expression Profiling; Q; R; 610; Infant; Prognosis; Risk Assessment; Progression-Free Survival; 004; 3. Good health; Gene Expression Regulation, Neoplastic; Neuroblastoma; 03 medical and health sciences; Deep Learning; Medicine; Humans; Female; Neoplasm Recurrence, Local; Child; Algorithms; Research Article

We introduce the CDRP (Concatenated Diagnostic-Relapse Prognostic) architecture for multi-task deep learning that incorporates a clinical algorithm, e.g., risk stratification schema to improve prognostic profiling. present first application survival prediction in High-Risk (HR) Neuroblastoma from transcriptomics data, task studies MAQC consortium have shown remain hardest among multiple diagnostic and endpoints predictable same dataset. To obtain more accurate needed appropriate treatment strategies, combines component (CDRP-A) synthesizing second (CDRP-N) dedicated one or tasks. The approach leverages advent of semi-supervised structures can flexibly integrate multimodal data internally create processing paths. CDRP-A is an autoencoder trained on gene expression HR/non-HR by Children's Oncology Group, obtaining 64-node representation bottleneck layer. CDRP-N classifier two endpoints, i.e., Event-Free Survival (EFS) Overall (OS). provides HR embedding input shared layer, which branches depart model EFS OS, respectively. control selection bias, evaluated using Data Analysis Protocol (DAP) developed within initiative. was applied Illumina RNA-Seq 498 patients (HR: 176) SEQC study (12,464 Entrez genes) Affymetrix Human Exon Array profiles (17,450 247 primary TARGET NBL cohort. On patients, achieves Matthews Correlation Coefficient (MCC) 0.38 MCC = 0.19 OS external validation, improving over published models. show indeed parametrically associated increasing severity be used better stratify patients' survival.