Susceptibility History Lab
Live research models

Susceptibility History Lab

New domestic models · CPU on the H200 host · separate from the MIMIC archive

This page sends the chosen input to the H200 server for a new model run. It is separate from the archived-result explorer. Use fictional or public research inputs; do not enter identifying patient information. The browser encrypts each input for the H200 server. The relay stores ciphertext only; plaintext is processed in H200 memory. Pending ciphertext expires after20 minutes and completed ciphertext after15 minutes. Cleanup runs on relay requests and the worker heartbeat.

Research demonstration. A successful run does not establish clinical reliability. The inference API has a fixed HTTPS address. The H200 worker reconnects outbound and resumes after a process failure; if the host is offline, inference is unavailable and the page reports it.

Connecting to the H200 service…

Run a fresh input

This workspace uses separate fitted domestic BAS-G/BSI-S model sets. It is a separate retrospective record-date proxy study, not a restored MIMIC model. CatBoost runs on the CPU of the H200 server. Prior result availability, laboratory timing, units and isolate linkage remain unverified.

These examples are fictional. The supplied culture.date may denote collection OR order date; the precise meaning is unverified. Legacy collection_date keys mean record dates. The two sensitivity variants validate the legacy clinical schema but exclude all18 current clinical predictors; their training cohort still reflects the original adult and clinical-snapshot-consistency selection. Neither gap guarantees prior report availability. Clinical AST is aminotransferase; history AST is the recorded S/I/R category. Only supported organism–drug cells are accepted. Missing clinical values remain missing. Same-day/future history is excluded; latest-date missing or conflicting results remain unresolved. Probabilities concern recorded S versus I/R; they are not prescribing advice.

Computational path

Browser encryption → bounded ciphertext relay → H200 decryption → fitted model → frozen Python measurement or correction → new response and model identity.

Model weights may stay loaded between calls. Returned inference results are not cached. B/C use the research GPU lock; the keeper yields during their model execution. A uses the host CPU. Failed calls show their failure; archived outputs are not substituted.