Susceptibility History Lab
Separate the value of historical information from the complexity of its representation.
Runs in your browser. No account, GPU or model API required.
Abstract
Historical susceptibility information, representation complexity and coverage in a retrospective benchmark
Historical antimicrobial susceptibility results may improve prediction, but this does not establish the value of a more complex history representation. We examined information content, representation complexity and history coverage in an existing MIMIC-IV benchmark. Patient-separated modelling and training-only calibration used 3,137 observed organism–drug pairs from 948 patients, including 214 evaluation patients. Expanded history had a patient-weighted Brier score of 0.1419, compared with 0.1619 for clinical predictors alone (difference -0.0200; 95% exploratory interval [-0.0357, -0.0040]). Compact history had a score of 0.1394. Expanded minus compact history was +0.0025 ([-0.0020, +0.0070]), and none of three stronger-control comparisons established superiority after Holm adjustment. Usable history was available for 50 of 214 evaluation patients; the descriptive improvement was concentrated in this group. These findings support a historical-information signal relative to the selected clinical baseline, without establishing additional value from representation expansion. Prior outcome exposure, selectively observed susceptibility results and uncertain clinical availability restrict interpretation to this retrospective benchmark.
What can this artifact reproduce?
The benchmark contains 948 patients and 3,137 observed organism–drug pairs. The frozen evaluation contains 214 patients and 716 pairs. The public artifact recomputes four aggregate contrasts and the three-test Holm family, and checks fictional history examples against the original extraction functions.
L_patient = mean over patients [mean over tested pairs (p − y)²] Δ = expanded-history Brier − comparator Brier
Patient-level inputs and fitted clinical predictors are unavailable here. The original intervals and raw p values are retained inputs; they are not recomputed from the aggregate table. This is an audit of the saved benchmark, not a patient prediction service.
Replay the fixed comparisons.
Negative differences favour expanded history. Confidence intervals are saved, pointwise and exploratory.
Fictional history-resolution lab
These records are fabricated. Select an edge case, inspect or edit its JSON, then run the same timing and conflict rules.
Why the patient is the analysis unit
Each inner array is one fictional patient. Patients contribute equal total weight even when pair counts differ.
From the page to the code.
The download includes this website, numerical samples, research reference code, provenance and executable checks. The examples are usable without access to a hospital or private account.
unzip history-lab.zip cd research-artifact python reproduce.py # Browse the static website locally: python -m http.server 8766 --directory dist
Download the complete installation package · Inspect provenance
Saved results are computational evidence. This artifact does not establish independent clinical validation, medical suitability or acceptance of a paper.