Visit my website · Explore the pipelines · Google Scholar · ORCID
I'm Tianyi Yu. I'm drawn to the stories hidden in clinical notes, the subtle signals in images, and the difficult questions behind a prediction.
I want to build AI that earns people's trust. My research explores how models behave when data are incomplete, populations change, and a good benchmark score meets the complexity of care. Here I share the code, questions and lessons along the way.
| 01 / Clinical NLP · 7 projects Biomedical language understanding, sentence classification and cross-corpus evaluation. |
02 / Computer Vision & Biomedical Imaging · 6 projects Skin-lesion modelling, image classification and selective referral. |
| 03 / Trustworthy AI · 9 projects Calibration, uncertainty, robustness and claims matched to evidence. |
04 / Prediction Modelling · 10 projects Clinical risk estimation, temporal validation and transportability. |
| 05 / Health Informatics · 10 projects EHR analysis, clinical data pipelines and population health surveillance. |
06 / AI for Health · 10 projects Reproducible machine learning for health research, with explicit evaluation and data-access boundaries. |
52 research projects across six connected directions. Here are the latest additions—each with its own code, pipeline guide and scientific cover.
 Does an augmentation gain remain meaningful when absolute predictive performance falls under observation loss? Explore the code → |  Do missingness stress tests change source-only model selection, or do competing rules choose the same candidate? Explore the code → |
 What can objective sleep architecture reveal without turning physiological disruption into a psychiatric label? Explore the code → |  What changes when an AKI progression model moves between clinical databases and requires local updating? Explore the code → |
Biomedical sentence classification examined through calibration and perturbation consistency. CLINICAL NLP · TRUSTWORTHY AI Explore pipeline → | Skin-lesion modelling with lesion-level splits, calibration and risk–coverage evaluation. COMPUTER VISION · BIOMEDICAL IMAGING Explore pipeline → |
ARDS and AHRF prediction across time and ICU databases, with explicit site-shift audits. PREDICTION MODELLING · HEALTH INFORMATICS Explore pipeline → | Next-week escalation modelling with rolling-origin validation and cluster-aware uncertainty. AI FOR HEALTH · POPULATION SURVEILLANCE Explore pipeline → |
Cross-organ gene-set classification under changing normalization and study weights. COMPUTATIONAL BIOLOGY · OPEN-DATA REANALYSIS Explore pipeline → | Word-level eye-tracking analysis with lexical features, grouped evaluation and a reusable CLI. COGNITIVE DATA SCIENCE · PYTHON PACKAGE Explore pipeline → |
Define the question. Make the endpoint, population and information available at prediction time explicit.
Test the transfer. Separate model development from temporal, site and cross-corpus evaluation.
Keep the evidence. Preserve uncertainty, negative findings, provenance and the limits of each experiment.
The full pipeline index includes data-access requirements and scope notes. These repositories host research software and reproduction guides; clinical projects are research-only.
Statistical care. Reproducible code. Claims matched to evidence.