Chenyu Wang (王晨瑜)

Guideline-as-Oracle: Ophthalmic Telephone Triage · First-author manuscript under review at AAAI

• Data & training: Developed a guideline-to-dialogue pipeline to address the cost of annotating multi-turn clinical conversations, converting a 70-row ophthalmic rule table into 3,000 training dialogues with rule-derived urgency labels. Performed full-parameter SFT of a 9B model using class resampling and nurse-role token supervision, and investigated Outcome GRPO as an exploratory post-training approach.
• Evaluation: Relative to the base 9B model, improved agreement from 61.7% to 74.1% and emergency recall from 9.5% to 69.0% on 201 internal rule-reference cases; checked robustness with a second training seed and patient simulator.

This reference also informed development and model selection, so these results measure internal policy conformance rather than held-out clinical generalization.

Guideline-as-Oracle paper ↗

Retrieval-Augmented mCRC MDT Decision Agent · Ongoing

• Data: Investigating how longitudinal records and retrieved cases support prediction of five recorded MDT management actions for metastatic colorectal cancer. Built a single-center data pipeline from 9,779 encounters / 4,983 patient records spanning 2013–2025; integrity screening retained 8,274 encounters / 4,077 patients. Restricted inputs to information available before each target meeting.
• Evaluation design: Froze a test set of 1,247 encounters from 610 patients, with no patient overlap across splits. Compared classical machine learning, retrieval, and language-model approaches using a common evaluation protocol.
• Method & results: Combined longitudinal history, similar-case retrieval, and XGBoost probability priors in a patient-time-aligned language-model prompt. Achieved 64.1% accuracy and 0.527 macro-F1, compared with 57.3% and 0.453 for the strongest classical baseline: gains of 6.8 percentage points in accuracy and 0.074 in macro-F1.
• Longitudinal information: For encounters with prior patient history, accuracy increased from 55.4% to 66.1% relative to the classical model.

Current results measure retrospective agreement with recorded MDT actions; prospective clinical benefit remains untested.

Ongoing

• Audit framework: Developing dependency-closure attribution and four-level evidence scoring across 12 security domains for agent software components. Coverage includes 207 repositories, 12,830 components, and 11,124 entry points. The method distinguishes shared-dependency findings from risks specific to individual components.
• Validation & reproducibility: Combined AST parsing with intra-procedural taint analysis to reduce items requiring manual review from 4,794 to 153; the remaining 153 represent 1.4% of all 11,124 entry points. Consolidated findings into 24 code defects. Two clean-environment reruns produced byte-identical results with fixed source and parser versions.

Graduate researcher · SIMIT, UCAS · First-author manuscript under review at IEEE Sensors Journal

• Method: Developed a single-stage unsupervised adaptation framework for temporal drift and cross-device distribution shift, combining hierarchical MK-MMD, CORAL, dual-level center regularization, and multi-head classification; also studied a relative-response front end and the conditional adversarial variant UnifiedGas-C.
• Results: On UCI Drift, obtained 81.10% / 84.36% mean accuracy for fixed-source / sequential transfer using retrospective target-label-based best-epoch selection. On the full-cycle Twin Arrays track, fixed-source / adjacent-board accuracy was 100.00% / 100.00% for UnifiedGas-C and 100.00% / 99.94% for CDAN under the same frontend-matched protocol.