Shipped · Dissertation
iamtanay/fedacuityFedAcuity
Master's dissertation in AI & ML. Federated learning for Long-Term Care facilities — predicting skilled nursing needs across Memory Care, Skilled Nursing, and Independent Living sites without ever moving patient data off-premises.
The core claim: care-type-aware federated aggregation outperforms a single global model — most decisively on fairness across care types. The dissertation is complete and the paper is written and bundled for arXiv.
Notes
Build log in reverse chronological order.
The statistical test I'd been reporting was meaningless
The third contribution, an honest fairness story, and throwing out a p-value that never had a sample behind it. The last FedAcuity entry.
The gap shrank from 13 points to 1.4, and that's the good news
Fixing an evaluation that was rigged in my own favor — a two-care-type held-out set, genuine multi-client clustering, and real MIMIC-IV access at last.
I audited my own results and they didn't survive
The midsem numbers looked great. Then I ran a research audit on my own pipeline and found six bugs holding the results up.
Three bugs and a framework swap later
Debugging Flower, replacing Ray, and what the numbers finally said when five FL strategies ran end-to-end
2325 lines before lunch
The FedAcuity scaffold — five FL strategies, a CTGAN synthetic data pipeline, and an IEEE paper skeleton, all built from a research question and a config file