Towards artificial intelligence-based disease prediction algorithms that comprehensively leverage and continuously...

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Clinical Epidemiology

Background:

• Real-world clinical data, such as electronic health records (EHR), contain vast information for disease risk prediction.
• Traditional methods often rely on curated datasets and pre-defined risk factors, limiting adaptability and scope.
• Developing dynamic, automated disease risk calculators from raw clinical data remains a challenge.

Purpose of the Study:

• To present a proof-of-concept for a generalizable, automated strategy (the full algorithm) to estimate disease risk using real-world clinical tabular data.
• To develop a self-adaptive prediction system that evolves with newly collected data.
• To demonstrate the feasibility of harnessing comprehensive real-world data for accurate disease risk prediction without extensive a-priori curation.

Main Methods:

• The full algorithm integrates statistical methods and artificial intelligence to parse EHR data, identify predictor variables, structure data into time-series, and train a neural network prediction model.
• The system is designed to be self-adaptive, automatically updating the prediction mechanism as new data become available.
• A pseudo-prospective validation was conducted using EHR data to estimate the 12-month risk of initial stroke or myocardial infarction.

Main Results:

• The algorithm achieved Area Under the Receiver Operating Characteristic Curve (AUROC) values ranging from 0.830 to 0.909 for predicting stroke or myocardial infarction risk, with an improving trend over time.
• Model precision, indicated by odds ratios, showed strong performance for high-risk patient groups (1-100 and 101-200), with values ranging from 7.2 to 48.1 and also demonstrating improving trends.
• The validation involved 558,105 patients and 3,424,060 patient-months of data from April 2015 to September 2023.

Conclusions:

• The proposed automated strategy is a feasible approach for developing high-performing, self-adaptive disease risk calculators using real-world clinical data.
• The algorithm effectively integrates diverse data types from EHRs to capture complex disease dynamics.
• This approach offers a powerful alternative to traditional static models, enabling more accurate and responsive risk estimation.