# PRISM-GI model card

Partial-domain-adjusted Rank-Integrated Survival Model for Gastrointestinal cancers.

## Intended use

Exploratory bulk-expression cohort risk stratification in colorectal, gastric and pancreatic cancer. Not for clinical diagnosis, treatment, screening or individual patient management. The browser output is a relative log-risk score, not a calibrated survival probability. The clinical nomogram is not implemented here.

## Frozen artifact

- Model: previously selected DeepSurv v1, not refitted for the website.
- Architecture: 15 inputs, 16 tanh hidden units, one bias-free output. Dropout is disabled during inference.
- Lambda: 0.5. Network seed: 20261103. CV seed: 20260903.
- Original weights SHA256: b13b6876bb25dadf54801e10286ccd124c4d2d7a013e5c3d10be802846ebae67.
- The model.json file contains the exported parameters, feature ordering and preprocessing constants.

## Exact inference

1. Accept processed expression for one source cohort and one cancer type. Rows are samples; columns are gene symbols. No gene alias resolution, probe aggregation or count normalization is performed in the browser.
2. Within each sample, calculate average-tie percentile ranks over the measured genes in the locked 158-gene background. Then select the fixed 15 model genes. Ranking only 15 genes is incorrect.
3. For each selected gene, calculate x_adjusted = x_rank - 0.5 * (mean_rank_of_uploaded_cohort - frozen_training_global_rank).
4. Apply the frozen feature means and population standard deviations; run the saved neural network in inference mode.
5. Report the network output as PRISM-GI score. Relative hazard = exp(score - uploaded_cohort_median_score). Group labels are descriptive comparisons to that median, not clinical cutoffs.

The local frozen artifact implements partial cohort-mean alignment. The website does not substitute a newly fitted GEO or TCGA subspace projector. The historical project description and user's cross-computer comparison do not independently establish mathematical equivalence between different operators.

## Input restrictions

All 15 panel genes are required. The website requires at least 150 of the 158 ranking-background genes; this is a conservative software guardrail, not an experimentally validated threshold. For an unmeasured non-panel gene, omit the entire column. All present background columns must be complete, finite and numeric. No silent sample filtering occurs.

Provide at least 2 samples; fewer than 20 triggers a small-cohort warning. No minimum sample size has been validated for deployment. Different uploaded cohort compositions lead to different adjustment means. Do not concatenate source cohorts, cancer types or arbitrary samples to score one patient. The cancer label does not change the shared network.

## Evidence and limitations

Training included 1,207 TCGA cases. GEO cohorts participated in exploratory model/configuration and seed selection. Those data are development-used rather than untouched independent clinical validation. Numerical agreement between JavaScript and frozen Python predictions does not imply clinical utility.

Public data downloads retain 207 pancreatic records after excluding the known overlapping GSE28735 source. Historical results based on a 249-record pancreatic pool are not republished as deduplicated performance claims. Source data may retain other unresolved limitations. Author processing, including GSE39582 RMA+ComBat, remains part of the input provenance.

## Privacy and downloads

The static browser application computes locally, without transmitting expression data to a prediction server, analytics or local storage. Hosting access logs are separate from prediction input. Input data is held in memory until cleared or the tab is reloaded/closed. Use pseudonymous sample identifiers. Public data downloads contain repository identifiers and OS annotations; follow original repository terms and cite original studies.

## Name

PRISM-GI is the name selected for this project. It is not affiliated with other PRISM-branded scores, tools or screening platforms.