Projects

AI as a tool for discovering new medical knowledge

Conventional medical research often begins with a predefined hypothesis: investigators select a disease and then design a test or model around it. My team takes a complementary, data-driven approach. We systematically link longitudinal ECGs, chest radiographs, laboratory tests, imaging studies, and clinical outcomes in the electronic health record to search for previously unrecognized examination-disease relationships. Here, AI is more than an automated reader. It serves as a digital microscope, revealing latent physiological states embedded in routine data but invisible to the unaided eye.
Our work shows that a single 12-lead ECG can reflect more than 50 physiological and disease phenotypes, including potassium imbalance, cardiac function, cardiopulmonary status, and near-term mortality risk. Our AI chest X-ray platform can identify more than 40 phenotypes and repurpose images acquired for other reasons for opportunistic screening, including osteoporosis. The value is not that AI repeats what clinicians already know how to see. It enables an existing examination to answer questions that conventional interpretation cannot, generating new digital biomarkers and testable medical hypotheses.

Selected references:

Lin C, Chen KC, Huang JW, Fang WH, Chang WC, Ko KH, Hsu YC, Lin CS, Lin SH, Tsai DJ* (2026, Jul) Transparent chest radiograph foundation model enables explainable human disease profiling. npj Digital Medicine, 9(1):674. [PDF]
Lin CS, Liu WT, Chen YH, Lin SH, Lin C* (2026, Jan) Artificial intelligence-enabled electrocardiography from scientific research to clinical application. EMBO Molecular Medicine, 18(1):22-40. [PDF]
Lin C, Chau T, Lin CS, Shang HS, Fang WH, Lee DJ, Lee CC, Tsai SH, Wang CH, Lin SH* (2022, Jan) Point-of-care artificial intelligence-enabled ECG for dyskalemia: a retrospective cohort analysis for accuracy and outcome prediction. npj Digital Medicine, 5(1):8. [PDF]

Establishing the real-world impact of AI through clinical trials

A newly discovered signal becomes clinically meaningful only when it changes care. Accuracy, regulatory authorization, and deployment alone do not prove that an AI system benefits patients. I therefore define the unit of clinical evidence as the complete indication-action pathway: who should be evaluated, when analysis should occur, which signal or threshold matters, how the result is delivered and interpreted, and what prespecified action follows.
In most applications, AI is an enabling diagnostic or risk-stratification component; the intervention is the downstream clinical response. The key question is not simply, “Does the AI work?” but “Does the AI-enabled care pathway improve patient outcomes?”
To answer that question, my team has built a digital clinical-trial infrastructure with more than 100,000 participants and more than 10 completed or ongoing pragmatic randomized trials embedded in routine care. Workflow-integrated AI-ECG alerts have reduced 90-day all-cause mortality, accelerated reperfusion for acute myocardial infarction, and prompted earlier treatment of potassium imbalance. In OPSCAN, AI reanalysis of existing chest radiographs identified people at high osteoporosis risk and directed them to confirmatory dual-energy X-ray absorptiometry (DXA), increasing new diagnoses and treatment initiation..

Selected references:

Lin CS, Liu WT, Tsai DJ, Lou YS, Chang CH, Lee CC, Fang WH, Wang CC, Chen YY, Lin WS, Cheng CC, Lee CC, Wang CH, Tsai CS, Lin SH, Lin C* (2024, May) AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial. Nature Medicine, 30(5):1461-1470. [PDF]
Lin C, Liu WT, Chang CH, Lee CC, Hsing SC, Fang WH, Tsai DJ, Chen KC, Lee CH, Cheng CC, Hung YJ, Lin SH, Tsai CS, Lin CS* (2024, Jun) Artificial Intelligence–Powered Rapid Identification of ST-Elevation Myocardial Infarction via Electrocardiogram (ARISE) — A Pragmatic Randomized Controlled Trial. NEJM AI, 1(7):AIoa2400190. [PDF]
Lin C, Tsai DJ, Wang CC, Chao YP, Huang JW, Lin CS, Fang WH* (2024, Jun) Osteoporotic Precise Screening Using Chest Radiography and Artificial Neural Network: The OPSCAN Randomized Controlled Trial. Radiology, 311(3):e231937. [PDF]
Lin C, Lin CS, Chen SJ, Tsai SH, Sung CC, Chen CC, Hsu YJ, Hung YJ, Lin SH* (2026, Jan) AI-enabled Electrocardiogram Alert for Potassium Imbalance Treatment: A Pragmatic Randomized Controlled Trial. Nature Communications, 17:159. [PDF]

Bringing hospital-grade diagnostic capabilities into communities and homes

The final step in clinical AI translation is to move validated capabilities beyond tertiary medical centers. Following technology transfer, authorization by the Taiwan Food and Drug Administration (TFDA), and U.S. FDA Breakthrough Device Designation, our AI-ECG platform has been deployed across more than 20 hospitals, emergency medical services, and over 60 public health centers, supporting cardiovascular screening for more than 20,000 residents in rural and underserved communities.
Home care presents a more demanding methodological challenge. We have assembled more than 50,000 Apple Watch and QOCA single-lead ECGs linked to electronic health records and developed models for mortality risk, ejection fraction, NT-proBNP, pulmonary artery pressure, potassium, hemoglobin, and other cardiopulmonary markers.
But moving a model into the home is not simply a matter of reducing a 12-lead ECG to one lead. Device characteristics, posture, signal quality, and use context can create domain shift and degrade performance. We therefore develop device-specific validation, calibration, and adaptation methods, while linking abnormal results back to confirmatory testing and treatment pathways. The goal is not to create isolated consumer diagnostics, but to make low-cost home sensing a safe and sustainable entry point into formal healthcare.

Selected references:

Lin C, Liu WT, Chen KC, Tsai DJ, Chang DW, Tsai TN, Cheng CC, Lin CY, Chen YH, Tsai CS, Lin SH, Lin CS* (2026, Jul) Bridging the gap from clinical to home ECG: quantifying and overcoming accuracy loss in AI-enabled single-lead ECG models. npj Digital Medicine, 9(1):510. [PDF]