Prediction and recommendation by machine learning through repetitive internal validation for hepatic veno-occlusive disease/sinusoidal obstruction syndrome and early death after allogeneic hematopoietic cell transplantation.

Using traditional statistical methods, we previously analyzed the risk factors and treatment outcomes of veno-occlusive disease/sinusoidal obstruction syndrome (VOD/SOS) after allogeneic hematopoietic cell transplantation. Within the same cohort, we applied machine learning to create prediction and recommendation models. We analyzed 2572 transplants using eXtreme Gradient Boosting (XGBoost) to predict post-transplant VOD/SOS and early death. Using the XGBoost and SHapley Additive exPlanations (SHAP), we found influential factors and devised recommendation models, which were internally verified by repetitive ten-fold cross-validation. SHAP values suggested that gender, busulfan dosage, age, forced expiratory volume, and Disease Risk Index were significant factors for VOD/SOS. The areas under the receiver operating characteristic curves and the areas under the precision-recall curve of the models were 0.740, 0.144 for all VOD/SOS, 0.793, 0.793 for severe to very severe VOD/SOS, and 0.746, 0.304 for early death. According to our single feature recommendation, following the busulfan dosage was the most effective for preventing VOD/SOS. The recommendation method for six adjustable feature sets was also validated, and a subgroup corresponding to five to six features showed significant preventive power for VOD/SOS and early death. Our personalized treatment set recommendation showed reproducibility in repetitive internal validation, but large external cohorts should prospectively validate our model.

Mechanism Design of Health Care Blockchain System Token Economy: Development Study Based on Simulated Real-World Scenarios.

BACKGROUND: Despite the fact that the adoption rate of electronic health records has increased dramatically among high-income nations, it is still difficult to properly disseminate personal health records. Token economy, through blockchain smart contracts, can better distribute personal health records by providing incentives to patients. However, there have been very few studies regarding the particular factors that should be considered when designing incentive mechanisms in blockchain. OBJECTIVE: The aim of this paper is to provide 2 new mathematical models of token economy in real-world scenarios on health care blockchain platforms. METHODS: First, roles were set for the health care blockchain platform and its token flow. Second, 2 scenarios were introduced: collecting life-log data for an incentive program at a life insurance company to motivate customers to exercise more and recruiting participants for clinical trials of anticancer drugs. In our 2 scenarios, we assumed that there were 3 stakeholders: participants, data recipients (companies), and data providers (health care organizations). We also assumed that the incentives are initially paid out to participants by data recipients, who are focused on minimizing economic and time costs by adapting mechanism design. This concept can be seen as a part of game theory, since the willingness-to-pay of data recipients is important in maintaining the blockchain token economy. In both scenarios, the recruiting company can change the expected recruitment time and number of participants. Suppose a company considers the recruitment time to be more important than the number of participants and rewards. In that case, the company can increase the time weight and adjust cost. When the reward parameter is fixed, the corresponding expected recruitment time can be obtained. Among the reward and time pairs, the pair that minimizes the company's cost was chosen. Finally, the optimized results were compared with the simulations and analyzed accordingly. RESULTS: To minimize the company's costs, reward-time pairs were first collected. It was observed that the expected recruitment time decreased as rewards grew, while the rewards decreased as time cost grew. Therefore, the cost was represented by a convex curve, which made it possible to obtain a minimum-an optimal point-for both scenarios. Through sensitivity analysis, we observed that, as the time weight increased, the optimized reward increased, while the optimized time decreased. Moreover, as the number of participants increased, the optimization reward and time also increased. CONCLUSIONS: In this study, we were able to model the incentive mechanism of blockchain based on a mechanism design that recruits participants through a health care blockchain platform. This study presents a basic approach to incentive modeling in personal health records, demonstrating how health care organizations and funding companies can motivate one another to join the platform.

Patient-Level Cancer Prediction Models From a Nationwide Patient Cohort: Model Development and Validation.

BACKGROUND: Nationwide population-based cohorts provide a new opportunity to build automated risk prediction models at the patient level, and claim data are one of the more useful resources to this end. To avoid unnecessary diagnostic intervention after cancer screening tests, patient-level prediction models should be developed. OBJECTIVE: We aimed to develop cancer prediction models using nationwide claim databases with machine learning algorithms, which are explainable and easily applicable in real-world environments. METHODS: As source data, we used the Korean National Insurance System Database. Every Korean in ≥40 years old undergoes a national health checkup every 2 years. We gathered all variables from the database including demographic information, basic laboratory values, anthropometric values, and previous medical history. We applied conventional logistic regression methods, light gradient boosting methods, neural networks, survival analysis, and one-class embedding classifier methods to effectively analyze high dimension data based on deep learning-based anomaly detection. Performance was measured with area under the curve and area under precision recall curve. We validated our models externally with a health checkup database from a tertiary hospital. RESULTS: The one-class embedding classifier model received the highest area under the curve scores with values of 0.868, 0.849, 0.798, 0.746, 0.800, 0.749, and 0.790 for liver, lung, colorectal, pancreatic, gastric, breast, and cervical cancers, respectively. For area under precision recall curve, the light gradient boosting models had the highest score with values of 0.383, 0.401, 0.387, 0.300, 0.385, 0.357, and 0.296 for liver, lung, colorectal, pancreatic, gastric, breast, and cervical cancers, respectively. CONCLUSIONS: Our results show that it is possible to easily develop applicable cancer prediction models with nationwide claim data using machine learning. The 7 models showed acceptable performances and explainability, and thus can be distributed easily in real-world environments.