Castration-Resistant Prostate Cancer Outcome Prediction Using Phased Long Short-Term Memory with Irregularly Sampled Serial Data

• Published in: Applied sciences Vol. 10; no. 6; p. 2000
• Main Authors: Park, Jihwan, Rho, Mi Jung, Moon, Hyong Woo, Lee, Ji Youl
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.03.2020
• Subjects: Algorithms; Cancer therapies; Castration; castration-resistant prostate cancer; Chemotherapy; Decision making; Deep learning; Electronic health records; Laboratories; Long short-term memory; Methods; Oversampling; Patients; phased long short-term memory; Physicians; Prediction models; Prostate cancer
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app10062000

It is particularly desirable to predict castration-resistant prostate cancer (CRPC) in prostate cancer (PCa) patients, and this study aims to predict patients’ likely outcomes to support physicians’ decision-making. Serial data is collected from 1592 PCa patients, and a phased long short-term memory (phased-LSTM) model with a special module called a “time-gate” is used to process the irregularly sampled data sets. A synthetic minority oversampling technique is used to overcome the data imbalance between two patient groups: those with and without CRPC treatment. The phased-LSTM model is able to predict the CRPC outcome with an accuracy of 88.6% (precision-recall: 91.6%) using 120 days of data or 94.8% (precision-recall: 96.9%) using 360 days of data. The validation loss converged slowly with 120 days of data and quickly with 360 days of data. In both cases, the prediction model takes four epochs to build. The overall CPRC outcome prediction model using irregularly sampled serial medical data is accurate and can be used to support physicians’ decision-making, which saves time compared to cumbersome serial data reviews. This study can be extended to realize clinically meaningful prediction models.