README.md

metadata

license: other
tags:
  - missing-data
  - imputation
  - benchmark
  - healthcare

Missing-Data Prediction Benchmark — results

Two experiments on 5 classification + 5 regression medical datasets under simulated missingness (MCAR/MAR/MNAR, rates 0.1–0.5, seeds 42/1/7), reported as mean [95% CI].

Headline

A cheap Active-Feature-Acquisition arm matches or beats a strong LLM baseline (gpt-5.6-terra) on accuracy while using 97.3–99.5% fewer tokens (cached; 76.2–95.3% under conservative per-step accounting).

Experiment 2 — LLM baseline vs AFA (headline)

dataset	mech	LLM acc	AFA acc	best AFA	LLM tokens	S2 cached	S2 uncached	↓ cached	↓ uncached
breast_cancer	MAR	0.915	0.965	smart	463,109	4,835	110,018	99.0%	76.2%
breast_cancer	MCAR	0.924	0.98	mim_smim	414,789	4,515	19,485	98.9%	95.3%
breast_cancer	MNAR	0.933	0.962	xgboost	480,151	4,835	110,615	99.0%	77.0%
gallstone	MAR	0.49	0.771	random_forest	1,103,203	6,043	60,795	99.5%	94.5%
gallstone	MCAR	0.495	0.755	mim_smim	830,642	6,043	43,630	99.3%	94.7%
gallstone	MNAR	0.49	0.76	random_forest	1,289,357	6,043	76,806	99.5%	94.0%
heart_disease	MAR	0.7	0.839	smart	258,031	2,068	25,029	99.2%	90.3%
heart_disease	MCAR	0.694	0.856	mim_smim	208,505	2,068	11,775	99.0%	94.4%
heart_disease	MNAR	0.694	0.833	smart	198,665	2,068	18,187	99.0%	90.8%
parkinsons	MAR	0.761	0.897	hist_gb	159,996	3,368	19,475	97.9%	87.8%
parkinsons	MCAR	0.744	0.872	hist_gb	234,958	3,211	22,513	98.6%	90.4%
parkinsons	MNAR	0.752	0.897	lightgbm	113,330	3,097	9,260	97.3%	91.8%

How to read (no download needed)

Every CSV opens in HuggingFace's Dataset Viewer (sortable, in-browser):

• experiment2_acquisition/acquisition_ab_report.csv — the table above.
• experiment1_prediction/prediction_classification_long.csv / prediction_regression_long.csv — one tidy row per (method, dataset, mechanism, rate, seed) with all metrics + status. This is the clean, flat view of every run (no folder-diving).
• experiment1_prediction/classification_mean_ci.csv / regression_mean_ci.csv — the same, collapsed to mean [95% CI] over seeds (blank cells carry a Status).
• experiment2_acquisition/curves_long.csv — per-budget line-plot data.
• DATA_DICTIONARY.md — column definitions and status meanings.
• raw_prediction/ — per-cell raw Exp-1 metric JSON (in-scope datasets, canonical layout, see raw_prediction/README_LATEST.md).
• raw_acquisition/ — per-cell raw Exp-2 AFA traces + LLM-baseline outputs (no model weights).

Data-use / MIMIC-III notice

One regression dataset (mimic_lengthofstay) is derived from MIMIC-III, credentialed under the PhysioNet Data Use Agreement. This repository contains only aggregate derived metrics — no patient-level records, no raw MIMIC data, and no models trained on MIMIC.

Generated 2026-08-11 by scripts/reporting/prepare_hf_upload.py.