# MIMIC-IV

ECG foundation models are usually compared on in-domain interpretation tasks, leaving their transfer to longitudinal prognosis across healthcare systems uncertain. PACE-ECG treats two recordings from the same patient within 60 days as a contrastive positive pair and adds waveform- and report-derived multitask supervision, with an encoder of 2.21 million parameters. On MIMIC-IV and two external HEEDB cohorts, its frozen representations predict post-myocardial-infarction all-cause mortality and incident heart failure with higher AUROC than DeepECG-SSL, with the largest gains in external heart failure.

ECG foundation models are usually compared on in-domain interpretation tasks, leaving their transfer to longitudinal prognosis across healthcare systems uncertain. PACE-ECG treats two recordings from the same patient within 60 days as a contrastive positive pair and adds waveform- and report-derived multitask supervision, with an encoder of 2.21 million parameters. On MIMIC-IV and two external HEEDB cohorts, its frozen representations predict post-myocardial-infarction all-cause mortality and incident heart failure with higher AUROC than DeepECG-SSL, with the largest gains in external heart failure.

ICU resource planning has long relied on compact state-transition models hand-specified for a single policy question, so a new question means redefining states. A patient world model trained on MIMIC-IV and then frozen encodes irregular electronic health records as state-event sequences, returns predictions at any interval the simulator requests, and connects through scenario-specific mappings to shared bed and staff queues. Across admission, care-dispatch and discharge simulations it agrees with task-fitted Markov and semi-Markov comparators on directional responses to capacity and dispatch policies, and reproduces the dispatch gains obtained by replaying observed trajectories.

ICU resource planning has long relied on compact state-transition models hand-specified for a single policy question, so a new question means redefining states. A patient world model trained on MIMIC-IV and then frozen encodes irregular electronic health records as state-event sequences, returns predictions at any interval the simulator requests, and connects through scenario-specific mappings to shared bed and staff queues. Across admission, care-dispatch and discharge simulations it agrees with task-fitted Markov and semi-Markov comparators on directional responses to capacity and dispatch policies, and reproduces the dispatch gains obtained by replaying observed trajectories.

Synthetic electronic health records have long traded clinical fidelity against formal privacy guarantees. SynthEHR-DP combines a mixed-type encoder, a temporal transformer encoder and a conditional variational autoencoder, trained under differentially private stochastic gradient descent with Rényi privacy accounting, then applies a post-hoc clinical constraint layer that spends no privacy budget. On MIMIC-IV at a total budget of ε=8 and δ=10⁻⁵, membership inference AUC is 0.52, downstream TSTR AUROC is 0.786, 4.2% below the non-private TRTR oracle of 0.821, and the clinical validity score is 0.961.

Synthetic electronic health records have long traded clinical fidelity against formal privacy guarantees. SynthEHR-DP combines a mixed-type encoder, a temporal transformer encoder and a conditional variational autoencoder, trained under differentially private stochastic gradient descent with Rényi privacy accounting, then applies a post-hoc clinical constraint layer that spends no privacy budget. On MIMIC-IV at a total budget of ε=8 and δ=10⁻⁵, membership inference AUC is 0.52, downstream TSTR AUROC is 0.786, 4.2% below the non-private TRTR oracle of 0.821, and the clinical validity score is 0.961.

Sepsis coagulopathy screening still rests on a single admission lab set, with platelets, INR and SOFA combined into one cross-sectional threshold. Among 3,138 sepsis patients in MIMIC-IV, clustering the joint 72-hour trajectories of platelet count, INR and PTT yielded a high-risk subphenotype with 27.8% mortality, of whom the admission SIC score flagged only 18.1%. The gradient reproduced in an independent refit of 12,828 eICU-CRD patients, where 86.3% of the high group remained SIC-negative.

Sepsis coagulopathy screening still rests on a single admission lab set, with platelets, INR and SOFA combined into one cross-sectional threshold. Among 3,138 sepsis patients in MIMIC-IV, clustering the joint 72-hour trajectories of platelet count, INR and PTT yielded a high-risk subphenotype with 27.8% mortality, of whom the admission SIC score flagged only 18.1%. The gradient reproduced in an independent refit of 12,828 eICU-CRD patients, where 86.3% of the high group remained SIC-negative.

A patient-safety reasoning system combines a knowledge graph, retrieval-augmented generation and a LLaMA model to extract safety concepts from MIMIC-III and MIMIC-IV records and produce traceable explanations. The complete system reached an F1 of 0.89 for concept extraction; removing the knowledge graph lowered F1 by 0.047, AUROC by 0.084 and AUPRC by 0.110, all significant. Removing retrieval or LoRA adaptation gave smaller, non-significant drops, and retrieval relevance was confirmed only qualitatively by experts.

A patient-safety reasoning system combines a knowledge graph, retrieval-augmented generation and a LLaMA model to extract safety concepts from MIMIC-III and MIMIC-IV records and produce traceable explanations. The complete system reached an F1 of 0.89 for concept extraction; removing the knowledge graph lowered F1 by 0.047, AUROC by 0.084 and AUPRC by 0.110, all significant. Removing retrieval or LoRA adaptation gave smaller, non-significant drops, and retrieval relevance was confirmed only qualitatively by experts.

Observational causal estimates gain a continuous reporting channel: the Bias Attribution Fraction (BAF) writes the share of an observed log relative risk attributable to systematic bias as a number between 0 and 1, with a 95% credible interval. Across 1,920,000 simulated repetitions, the calibrated P-value alone declared 36.02% of truly bias-dominated repetitions usable, adding the BAF point estimate cut that to 16.48%, and requiring the whole credible interval below 0.5 cut it to 10.16%. In a MIMIC IV beta-blocker target trial emulation, the 16.9% one-year mortality reduction associated with reaching target dose disappeared after calibration, with BAF 0.875 and a verdict of not usable as effect evidence.

Observational causal estimates gain a continuous reporting channel: the Bias Attribution Fraction (BAF) writes the share of an observed log relative risk attributable to systematic bias as a number between 0 and 1, with a 95% credible interval. Across 1,920,000 simulated repetitions, the calibrated P-value alone declared 36.02% of truly bias-dominated repetitions usable, adding the BAF point estimate cut that to 16.48%, and requiring the whole credible interval below 0.5 cut it to 10.16%. In a MIMIC IV beta-blocker target trial emulation, the 16.9% one-year mortality reduction associated with reaching target dose disappeared after calibration, with BAF 0.875 and a verdict of not usable as effect evidence.