OBJECTIVES:

Classical Markov models often use fixed values as transition probabilities between states, which limits their clinical applications in real-world due to ignoring the impact of individualized characteristics on disease progression. In this study, a framework for cost and utility prediction in patients with GI tract tumors was developed based on Markov and deep-learning algorithms.

METHODS:

In this study, a Markov model with a cycle period of 6 months and containing initial, progressive and death states was set, and a prediction model for the prediction of total cost and transfer probability between states was established based on Deep Neural Networks and CatBoost using electronic medical record data (EMR) from 456 patients with 1,956 follow-up visits. Health utilities, assessed by EQ-5D-5L, were obtained from 530 patients in a cross-sectional process, and a model for predicting health utility values was developed based on EMR and CatBoost. The results of the above three prediction models were used as parameters for the Markov model to establish a Deep Learning-Markov model framework, and a web application platform was developed for use by interested parties.

RESULTS:

The training rounds of the transfer probability, cost, and health utility value prediction models were 10,000, 1,000, and 1,000, and RMSEs were 0.35, 1711.7625, and 0.0934, respectively. The training loss of each model is continuously reduced and successfully converged. The deep learning-Markov model framework and application platform developed based on the models can output the life year, QALY and ICER for individual patient under different treatment regimens.

CONCLUSIONS:

In this study, we developed a model that can be used to individually predict long-term cost-utility data for patients, which can provide individualized cost and utility values to help the selecting high-value treatment options for oncology patients and has potential clinical application.