PRISM

Cross-institutional deployment of medical models requires supervised external adaptation and cannot rely on zero-shot direct transfer.

On the external MC-MED cohort, the directly transferred source-only PRISM achieved only 0.7841 AUROC, falling below the locally trained gradient boosting model's 0.8227.

After supervised fine-tuning, PRISM improved to 0.8611 AUROC on the external MC-MED cohort, surpassing the locally trained traditional model.

Structure-aware multimodal integration outperforms the approach of serializing everything into text.

PRISM (Population-Retrieved Inference for Structured Multimodal Prognosis) is a retrieval-augmented multimodal language model used to integrate heterogeneous clinical evidence for clinical prognostic prediction.

PRISM performs structured processing at three key stages: applying independent numerical tokenization to each numerical feature, using recency-weighted aggregation for longitudinal ECGs, and constructing a bimodal population memory bank.

PRISM's retrieval operates in representation space, where the model receives mean-pooled embeddings from visit records of similar training cohorts rather than individual traceable patient records.

In ablation experiments, replacing numerical EHR encoding with text serialization resulted in a 3.66 percentage point drop in AUROC, the largest impact among all interventions tested.

On the MDS-ED benchmark comprising 1,443 tasks, PRISM achieved a diagnostic AUROC of 86.2%, deterioration prediction of 93.1%, ICU admission prediction of 91.8%, and mortality prediction of 92.8%.

PRISM's high scores utilized a 96-hour pre-arrival time window, a condition not available to the comparison baseline models.

In a fair comparison using only the baseline window input, PRISM's AUROC was 0.8895, broadly on par with UniPACT's 0.8937 and MDS-ED's 0.8889.

A compact fusion model using the same five evidence streams and frozen encoders achieved only 0.7652 AUROC under matched inputs, while PRISM reached 0.9092 under the same settings.

Ablation experiments showed that removing pre-arrival medical history caused AUROC to drop by 1.97 points, and zeroing out retrieval caused AUROC to drop by 0.99 points.

True retrieval outperformed random neighbor retrieval by 0.97 points on the macro-average, but a paired task-level sign test did not establish a consistent advantage (p=0.067).

The authors performed cross-platform replication using the independent data platform PRISM, achieving 82% directional consistency, but with weak rank correlation (rho=0.25).

PRISM is a high-throughput drug screening platform developed by the Broad Institute that barcodes cell lines and uses mixed-culture co-culture to simultaneously measure the response of multiple cell lines to the same drug.

The U.S. National Security Agency (NSA) collected user data directly from the servers of tech giants such as Google, Microsoft, and Apple through programs including PRISM.