Projects

Research question

Among patients with non-small cell lung cancer (NSCLC) who received neoadjuvant immunochemotherapy followed by curative surgery, investigate the predictive value of primary-tumor and lymph-node features from pre- and post-treatment chest CT for patient-level postoperative pathological nodal status (ypN).

Work to date

• Audit image–mask geometry, feature extraction, and cohort manifests to trace information loss in preprocessing and local representations.
• Use fixed patient-level splits, training-only fitting, and leave-one-center-out (LOCO) validation to compare features and models and examine dependence on center differences.
• Compare Logistic baselines, single-factor fixes, and candidates including TabPFN; assess paired uncertainty, calibration, and cross-center stability before retaining changes.

Current stage

The study is ongoing, focusing on reliable baselines, identifying information loss, and improving representations. Current candidates do not yet support stable adoption; no final model has been selected.

Problem

Simulate Hong Kong unsecured installment lending and turn ongoing customer, application, disbursement, and repayment events into consistently defined, traceable daily business metrics.

Methods

• Organize raw, staging, mart, and audit layers with Python and PostgreSQL, implementing incremental loading, event versioning, and reconciliation.
• Schedule daily processing and correction-window rebuilds locally, publishing only aggregate metrics to the cloud for a five-page Streamlit dashboard.
• Verify rerun consistency, transaction rollback, and access permissions, with automated tests in GitHub Actions.

Results

The verified project snapshot covers 248 days, 4,665 disbursed loans, and HK$94.7M in principal, with 19 passing tests. The dashboard covers the executive overview, credit funnel, portfolio and collections, origination vintages, and channel performance and data quality.

All figures refer to synthetic data and a project acceptance snapshot, not real customers, actual lending volume, or the dashboard’s current live values.

Problem

Do emails improve visits, purchases, and spending? Across no email, men’s email, and women’s email, does personalized action selection provide more offline value than uniform targeting?

Methods

• Conduct a post-hoc analysis of 64,000 records from the historical Hillstrom randomized experiment, including data quality checks, baseline balance diagnostics, and hypothesis tests.
• Restrict model and policy selection to training procedures using nested cross-validation, and evaluate three-action policies with augmented inverse probability weighting (AIPW).
• Assess stability across five prespecified random seeds, distinguishing overall email effects from the added value of personalization.

Results

Men’s email increased two-week spending per customer relative to no email by approximately US$0.77 (95% CI: US$0.49–1.05). This between-group effect is distinct from whether personalization outperforms uniform targeting; the current evaluation did not establish an advantage over the best uniform policy selected within training.

This is a reanalysis of a historical experiment, not a newly conducted online A/B test. A nonsignificant difference does not establish policy equivalence.

Problem

Functional data such as multilead ECGs are high-dimensional and strongly correlated. How can shared and function-specific risk components be learned under right-censored survival outcomes while retaining traceable interpretations?

Methods

• Extend multiscale filtration representations to the score and information geometry of the Cox partial likelihood; compare Fusion and Directional grouping and extract residual risk directions layer by layer.
• Save basis functions, recursive mappings, scaling parameters, and Cox coefficients to reconstruct fitted risk scores as lead-by-time loadings.
• Compare multiple split seeds, five-fold evaluation, and nested training sizes on CODE-15 twelve-lead ECG data, with numerical and structural regression tests of the core implementation.

Results

The public repository reports 154,744 ECG records and 7,888 deaths. At full training size, Directional and Fusion use an average of 36.48 and 25.02 components, respectively, compared with 72.00 for the non-sharing functional Cox model, with similar observed discrimination.

These figures come from the repository report; experiments were not rerun for this website update. Similar observed performance is not statistical equivalence. Comparisons across estimators and repetition counts are descriptive. Lead associations describe the fitted model, not physiological connections or causal relationships.

Problem

How can follow-up information support prognosis modeling and comparisons of sequence models and survival methods within a leakage-controlled evaluation pipeline?

Methods

• Integrate longitudinal records from 5,368 MSK-CHORD breast cancer patients into a Landmark and Transformer pipeline, predicting subsequent survival risk at multiple follow-up points with patient-level splits and training-only preprocessing.
• Train with a numerically stable Cox partial likelihood, conduct five-fold internal validation, and use SHAP to analyze feature contributions.
• Examine how predictions respond to changes in disease-progression variables through model sensitivity analysis.

Results

Five-fold internal validation yielded a C-index of 0.887 ± 0.009 (mean ± SD), measuring risk-ranking discrimination.

These are internal-validation results, not external clinical validation or deployment outcomes. Changes to disease-progression variables represent model sensitivity, not identified causal intervention effects.