Projects

Enhancing Cross-Population and Disease Prediction Through Construction of a General Demographic Foundation Model

Demographic information, including gender, age, and ethnicity, represents widely available clinically data that carries rich insights about individual patients. Despite its value, such information is often underutilized in medical AI models. This study had constructed a general demographic foundation model based on the National Health Insurance Research Database of Taiwan. Results demonstrate that incorporating only age and gender attributes, the model has significantly improved diagnostic prediction performance for osteoporosis (U.S. dataset, AUROC increased from 92% to 99.9%) and thyroid disease (Australian dataset, AUROC increased from 83% to 99.7%).

Enhancing Cross-Disease Prediction: Development of a General Laboratory Progression Foundation Model

Using innovative self-supervised learning techniques, we constructed a pretrained model based on laboratory trajectories of patients with diabetes and hypertension. This model was successfully transferred to the prediction of outcomes in patients undergoing percutaneous coronary intervention (PCI), a population too small to train deep learning models independently. The approach improved prediction accuracy for target vessel revascularization (65% to 92%), major adverse cardiovascular events (52% to 71%), cardiovascular death (60% to 75%), myocardial infarction (43% to 68%), stroke (57% to 67%), and major adverse cardiac and cerebrovascular events (55% to 76%). The method requires no manual labeling, effectively handles irregular and missing values, and relies on only six common laboratory tests, ensuring no additional clinical burden.

Multimodal Clinical Data Integration: Prognostic Prediction of Spinal Surgery for Low Back Pain and Sciatica

The outcomes of spinal surgery for severe low back pain and sciatica are difficult to predict, with few effective preoperative assessment tools available. This study integrates the western and eastern medicine perspectives, collecting multi-modality data, including structured medical records, preoperative plans and summaries (text data), and vowel pronunciation recordings (audio data). The constructed multi-modality fusion learning model achieved 81% predictive accuracy, higher than average manual prediction (61%), demonstrating the value of integrating diverse clinical information sources benefits in complex treatment decision.

Simplifying Diagnostic Procedure: Blood Test-Based Detection of Primary Aldosteronism

Primary aldosteronism shares symptoms with hypertension but requires entirely different treatments, making early and accurate diagnosis critical. However, the task requires medical imaging and tissue sampling through operation. This study identified genetic variation in the potassium channel KCNJ5 as a diagnostic marker and developed a machine learning model using only blood test data. The model achieved an accuracy of 86% in detecting KCNJ5 mutations, simplifying the diagnostic procedure and opening the opportunity for large-scale disease screening in the future.

Supporting Healthcare Resource Allocation: Predicting Patient Healthcare-Seeking Behavior with Deep Learning

Patients’ choices while accessing healthcare often rely on subjective factors such as personal preference, prior experience, peer recommendations, or hospital reputation. This results in imbalanced healthcare utilization, with overcrowding in medical centers and underuse of smaller hospitals. Using NHIRD data, we constructed deep learning models to predict patients’ choice of hospital distance and hospital level with 96.8% accuracy. These models enable better understanding and simulation of healthcare demand, providing valuable insights for optimizing resource allocation during health policy making.