TRACER, a time-to-event EHR foundation model for disease detection across health systems

Mara Pleasure; Ekaterina Redekop; Vedrana Ivezic; Rushikesh Kulkarni; Aaron Chin; William Speier; Corey W. Arnold

cs.CEOct 29, 2026

TRACER, a time-to-event EHR foundation model for disease detection across health systems

Authors: Mara Pleasure, Ekaterina Redekop, Vedrana Ivezic, Rushikesh Kulkarni, Aaron Chin, William Speier, Corey W. Arnold

Organizations: Biomedical AI Research Lab, University of California, Los Angeles, Los Angeles, CA, USA · Department of Radiology, University of California, Los Angeles, Los Angeles, CA, USA · Bioinformatics, University of California, Los Angeles, Los Angeles, CA, USA · Department of Bioengineering, University of California, Los Angeles, Los Angeles, CA, USA ·

Abstract

Foundation models trained on longitudinal electronic health records (EHRs) can predict clinical diagnoses years in the future, making them a potential tool for disease screening. We developed TRACER (Temporal Risk Assessment with Clinical EHR Representations), a decoder-based foundation model pretrained on data from over 4.5 million patients from four health systems, and evaluated it on 16 incident diseases at 1-, 2-, and 5-year horizons in more than 2 million patients from two health systems excluded from model development. In contrast to typical evaluations that focus on global metrics (e.g., AUROC), we measured performance at the low false-positive rates that are realistic for screening. We found that TRACER retained its discrimination in external health systems, and at 95% specificity, the patients it flagged were up to eight times more likely to be diagnosed within a year than unselected patients. Evaluating at this operating point also revealed what global metrics alone did not: detection and false-alert burden varied across diseases and health systems, even when discrimination appeared stable. Our results illustrate the potential of foundation models for screening and underscore the need for evaluation in health systems excluded from development to assess disease detection and false-alert burden at levels appropriate for intended clinical use.}

Figures & tables

Figure 1: Multi-center EHR foundation modeling and time-to-event prediction framework. (a) Cohort allocation across six University of California (UC) health systems. Patients with less than one year of observable EHR history were excluded. UC1-UC4 provided the pretraining data and the held-out in-domain patients used to train and evaluate disease-prediction heads; UC5 and UC6 were excluded from both and used only for external evaluation. (b) TRACER modeling pipeline. Each patient’s history is a temporally ordered sequence of clinical tokens processed by a causal decoder-only transformer pretrained to predict the next token. (c) Discrete-time labels. Each patient contributes one interval index to the DeepHit likelihood. (d) Evaluation at each horizon. Cases are patients diagnosed by the horizon and controls are patients followed event-free beyond it with inverse probability of censoring weights (IPCW) for earlier censoring. Metrics include cumulative/dynamic AUROC, the standardized partial AUROC over false-positive rates 0-10|% (shaded), and sensitivity, PPV, and case yield at cohort-specific thresholds targeting 90%, 95%, and 99% specificity (points). Uno’s IPCW C-index summarizes event-time ranking across all horizons.

Figure 2: TRACER Discrimination across health systems and prediction horizons (a) Uno’s IPCW C-index for TRACER in the in-domain test cohort (yellow circles) and the held-out UC5 (blue diamonds) and UC6 (red-orange triangles) cohorts, with grey connectors spanning the lowest to the highest value for each endpoint. The right-hand column lists the worst-site gap, calculated as the in-domain C-index minus the lower of the two external values. (b) TRACER cumulative/dynamic AUROC at the 52-, 104-, and 260-week horizons for each cohort, with point estimates printed in each cell.

Figure 3: Disease detection at low false-positive rates. Sensitivity was calculated for TRACER at cohort-specific thresholds selected to target the stated specificity, using the same censoring-aware case and control definitions as AUROC with IPCW. (a) Sensitivity (%) at 95% specificity for each endpoint at the 52-, 104-, and 260-week horizons in the in-domain, UC5, and UC6 cohorts with values printed in each cell. (b) Sensitivity at the 52-week horizon at threshold target 90% (open circles), 95% (filled circles), and 99% specificity (squares), one panel per cohort. (c) Sensitivity at 95% specificity and 52 weeks for chronic kidney disease (solid line) and heart failure (dashed line) across the three cohorts. Error bars are 95% confidence intervals from 1,000 patient-level bootstrap replicates. Endpoints in (a) and (b) are listed alphabetically. Marker colors and shapes denote the evaluation cohort (yellow circle = in-domain, blue diamond = UC5, red-orange triangle = UC6).

Figure 4: Positive predictive values and case yield at 95% specificity over 52 weeks. For each endpoint and cohort, the TRACER score threshold was selected within the evaluation cohort to attain nominal 95% specificity, and PPV and case yield were censoring adjusted estimates. (a) PPV against one-year baseline incidence for each endpoint and cohort on logarithmic axes. Grey segments join each endpoint’s in-domain point to its UC5 and UC6 points. The solid diagonal marks PPV equal to baseline incidence (no enrichment), and dashed lines mark 2-, 5-, and 10-fold enrichment. The seven endpoints discussed in the text are labeled. (b,c) Incident cases recovered per 10,000 score records at the 52-week horizon, calculated at 10,000 times the IPCW weighted number of true positives divided by the number of scored records. Endpoints with an in-domain yield of at least 20 cases per 10,000 are shown in (b) on a 0-300 scale, and the remaining endpoints in (c) on a 0-25 scale; within each panel, endpoints are ordered by in-domain yield. Error bars are 95% confidence intervals. Colors denote the evaluation cohort (yellow = in-domain, blue = UC5, red-orange = UC6).

Figure 5: Overall discrimination of TRACER and TRACER-Encoder across 16 disease endpoints and three health systems. Both pretrained models remained fixed while endpoint-specific DeepHit survival heads were trained on labeled in-domain patients and applied unchanged to the held-out UC5 and UC6 cohorts. Endpoints are ordered by TRACER in-domain C-index in both panels. (a) Uno’s IPCW C-index for TRACER (solid bars) and TRACER-Encoder (hatched bars) in the in-domain test cohort, UC5, and UC6. Error bars are model-specific 95% confidence intervals. (b) Paired difference in cumulative/dynamic AUROC at the 52-week horizon, calculated as TRACER minus TRACER-Encoder, for each endpoint and cohort. Positive values favor TRACER and negative values favor TRACER-Encoder. Error bars are 95% confidence intervals from 1,000 paired patient-level bootstrap replicates in which the same resampled patients were applied to both models; bars who intervals includes zero are drawn in outline only. Colors denote the evaluation cohort (yellow = in-domain, blue = UC5, red-orange = UC6.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Chronic disease endpoints
	Development domain			Held-out health systems
Disease endpoint	Training	Validation	Test	UC5	UC6
Hypertension	10.00	10.00	10.00	11.70	15.20
Hyperlipidemia	9.23	9.23	9.25	15.47	9.42
Depression	6.81	6.82	6.81	8.78	6.88
Chronic kidney disease	2.53	2.57	2.54	2.51	2.53

Appendix

Table S1: Observed endpoint-event frequency across development and held-out cohorts. Values are the percentage of eligible patients with an observed incident endpoint event during available follow-up in the development-domain downstream training, validation, and test cohorts and the held-out UC5 and UC6 cohorts. Rows are ordered by event frequency in the in-domain training cohort within each endpoint group. Cohort sizes varied by endpoint-specific eligibility criteria; ovarian cancer cohorts were restricted to female patients.

Chronic disease endpoints
	In-domain		UC5		UC6
Disease endpoint	Decoder	Encoder	Decoder	Encoder	Decoder	Encoder
Heart failure	$0.869\,{\pm}\,0.015$	$0.810\,{\pm}\,0.021$	$0.865\,{\pm}\,0.003$	$0.823\,{\pm}\,0.004$	$0.836\,{\pm}\,0.004$	$0.798\,{\pm}\,0.005$
Acute myocardial infarction	$0.842\,{\pm}\,0.029$	$0.775\,{\pm}\,0.046$	$0.830\,{\pm}\,0.005$	$0.776\,{\pm}\,0.007$	$0.862\,{\pm}\,0.013$	$0.798\,{\pm}\,0.020$
Lupus	$0.763\,{\pm}\,0.079$	$0.701\,{\pm}\,0.098$	$0.746\,{\pm}\,0.012$	$0.711\,{\pm}\,0.012$	$0.669\,{\pm}\,0.020$	$0.661\,{\pm}\,0.023$
Hypertension	$0.776\,{\pm}\,0.010$	$0.768\,{\pm}\,0.010$	$0.761\,{\pm}\,0.002$	$0.753\,{\pm}\,0.002$	$0.775\,{\pm}\,0.003$	$0.753\,{\pm}\,0.003$

Appendix

Table S2: Global discrimination across the in-domain and held-out health-system cohorts. Uno’s IPCW C-index is reported for TRACER and TRACER-Encoder using survival heads trained on the full development-domain downstream training cohort and applied unchanged to the in-domain, UC5, and UC6 evaluation cohorts. Values are shown as the point estimate

\pm

the half-width of the model-specific analytic 95% confidence interval.

Chronic disease endpoints
Disease	In-domain	UC5	UC6	Worst-site gap
Heart failure	$0.869\,\pm\,0.015$	$0.865\,\pm\,0.003$	$0.836\,\pm\,0.004$	$+0.033$
Acute myocardial infarction	$0.842\,\pm\,0.029$	$0.830\,\pm\,0.005$	$0.862\,\pm\,0.013$	$+0.013$
Lupus	$0.763\,\pm\,0.079$	$0.746\,\pm\,0.012$	$0.669\,\pm\,0.020$	$+0.095$
Hypertension	$0.776\,\pm\,0.010$	$0.761\,\pm\,0.002$	$0.775\,\pm\,0.003$	$+0.015$
Celiac disease	$0.624\,\pm\,0.091$	$0.617\,\pm\,0.012$	$0.651\,\pm\,0.021$	$+0.007$

Appendix

Table S5: TRACER overall discrimination across health systems. Uno’s IPCW C-index for the development-trained TRACER model, applied unchanged to the in-domain and two external evaluation cohorts. Values are point estimates

\pm

the half-width of the model-specific analytic 95% confidence interval. The reported worst-site gap is the in-domain C-index minus the lower of the UC5 and UC6 values.

Disease	52 weeks	104 weeks	260 weeks	$\Delta_{52\rightarrow 260}$
Small 52-to-260-week change under TRACER ( $\lvert\Delta\rvert\leq 0.025$ )
Lupus	0.778	0.807	0.795	$+0.017$
Atrial fibrillation	0.852	0.861	0.859	$+0.008$
Rheumatoid arthritis	0.786	0.778	0.784	$-0.002$
Heart failure	0.870	0.869	0.862	$-0.008$
Hyperlipidemia	0.785	0.773	0.776	$-0.009$

Appendix

Table S6: TRACER AUROC across prediction horizons in-domain. Time-dependent AUROC at the 52-, 104- and 260-week horizons, using disease-prediction heads trained on the full in-domain training cohort. The final column reproduces the reported change from 52 to 260 weeks.

		In-domain				UC5				UC6
Disease	Model	26 wk	52 wk	104 wk	260 wk	26 wk	52 wk	104 wk	260 wk	26 wk	52 wk	104 wk	260 wk
Chronic disease endpoints
Heart failure	D	—	0.870	0.869	0.862	—	0.868	0.869	0.877	—	0.850	0.844	0.836
	E	—	0.829	0.821	0.822	—	0.840	0.838	0.839	—	0.815	0.820	0.814
Acute myocardial infarction	D	—	0.871	0.892	0.843	—	0.854	0.844	0.830	—	0.858	0.863	0.849
	E	—	0.794	0.812	0.771	—	0.766	0.773	0.776	—	0.791	0.811	0.801

Appendix

Table S7: Time-dependent AUROC by prediction horizon in the in-domain, UC5, and UC6 evaluation cohorts. Point estimates of cumulative/dynamic IPCW AUROC for TRACER (D) and TRACER-Encoder (E), survival heads trained on the full in-domain training cohort and applied unchanged externally. Chronic endpoints have no 26-week evaluation. 95% bootstrap confidence intervals are in the accompanying data file.

Heart failure
	Direct patient selection			Censoring-adjusted outcomes
Cohort	Patients scored	Patients selected	Selected (%)	Sensitivity (%)	Not identified (%)	PPV (%)	Case yield per 10,000
In-domain	28,800	1,505	5.23	44.18	55.82	13.80	78.59
UC5	919,344	51,998	5.66	36.88	63.12	9.47	51.59
UC6	283,392	15,279	5.39	33.57	66.43	16.41	95.39
Pancreatic cancer

Appendix

Table S8: TRACER patient selection and disease detection at 52 weeks. Results at thresholds selected retrospectively within each evaluation cohort to target 95% specificity. The full development-trained disease-prediction heads were applied unchanged; only the cohort-specific evaluation thresholds differed. “Not identified” is the estimated percentage of future cases below the threshold (100% minus sensitivity), not the percentage of all scored patients.

Chronic disease endpoints
	In-domain			UC5			UC6
Disease	90%	95%	99%	90%	95%	99%	90%	95%	99%
Heart failure	63.7	44.2	15.1	56.5	36.9	11.3	51.2	33.6	10.8
Acute myocardial infarction	55.8	42.0	6.2	54.9	36.2	11.8	57.3	39.6	15.9
Lupus	45.8	26.0	6.4	44.5	32.3	15.0	26.4	19.0	7.9
Hypertension	41.4	25.6	7.1	36.0	21.3	5.5	40.1	25.2	7.4

Appendix

Table S9: TRACER sensitivity at fixed specificity at 52 weeks. Entries are estimated sensitivity (%) at cohort-specific thresholds targeting the specificity shown in the column headings. The same development-trained model was applied without local adaptation. Outcome estimates account for incomplete follow-up.

		In-domain			UC5			UC6
Disease	Horizon (wk)	90%	95%	99%	90%	95%	99%	90%	95%	99%
Chronic disease endpoints
Heart failure	52	63.7	44.2	15.1	56.5	36.9	11.3	51.2	33.6	10.8
	104	62.9	45.0	13.3	57.9	39.1	11.0	49.8	33.1	9.5
	260	60.4	42.4	14.5	60.3	43.1	15.4	48.4	32.1	9.9
Acute myocardial infarction	52	55.8	42.0	6.2	54.9	36.2	11.8	57.3	39.6	15.9

Appendix

Table S10: TRACER sensitivity at fixed specificity across all prediction horizons. Estimated sensitivity (%) at cohort-specific thresholds targeting 90%, 95%, and 99% specificity, for survival heads trained on the full in-domain training cohort and applied unchanged to UC5 and UC6. Chronic endpoints were evaluated at 52, 104, and 260 weeks; cancers additionally at 26 weeks. Estimates account for incomplete follow-up (IPCW); 95% bootstrap confidence intervals are provided in the accompanying data file.

Cohort	Scored records	Baseline incidence (%)	Sensitivity (%) (95% CI)	PPV (%) (95% CI)	Enrichment (fold)	Cases per 10,000 (95% CI)
Heart failure
In-domain	28,800	1.779	44.18 (39.46–50.63)	13.80 (11.86–15.95)	7.76	78.59 (70.19–90.07)
UC5	919,344	1.399	36.88 (36.02–37.94)	9.47 (9.19–9.76)	6.77	51.59 (50.40–53.07)
UC6	283,392	2.842	33.57 (32.45–35.14)	16.41 (15.85–17.14)	5.78	95.39 (92.23–99.86)
Acute myocardial infarction
In-domain	29,112	0.362	42.04 (31.15–54.81)	2.96 (2.00–4.04)	8.19	15.20 (11.26–19.82)

Appendix

Table S11: TRACER disease detection, PPV and fold enrichment at one year. Results at retrospectively selected, cohort-specific thresholds targeting 95% specificity, using the frozen decoder-based TRACER representations and disease-prediction heads trained on 100% of the in-domain downstream training cohort and applied unchanged externally. Sensitivity, PPV and case yield are shown with the existing 95% confidence intervals in parentheses. Baseline incidence is the matched IPCW-weighted one-year case pool divided by the scored-record denominator. Fold enrichment equals PPV divided by that baseline incidence.

	In-domain		UC5		UC6
Disease endpoint	D	E	D	E	D	E
Heart failure	$0.877\,/\,0.506$	$0.838\,/\,0.611$	$0.905\,/\,0.530$	$0.875\,/\,0.431$	$0.865\,/\,0.533$	$0.843\,/\,0.403$
Acute myocardial infarction	$0.843\,/\,0.669$	$0.819\,/\,0.734$	$0.875\,/\,0.588$	$0.837\,/\,0.353$	$0.827\,/\,0.581$	$0.775\,/\,0.473$
Lupus	$0.811\,/\,0.641$	$0.692\,/\,0.535$	$0.843\,/\,0.429$	$0.794\,/\,0.488$	$0.739\,/\,0.473$	$0.665\,/\,0.496$
Hypertension	$0.812\,/\,0.576$	$0.784\,/\,0.572$	$0.820\,/\,0.667$	$0.808\,/\,0.595$	$0.814\,/\,0.437$	$0.794\,/\,0.558$
Celiac disease	$0.713\,/\,0.560$	$0.637\,/\,0.556$	$0.728\,/\,0.443$	$0.708\,/\,0.491$	$0.748\,/\,0.455$	$0.669\,/\,0.477$

Appendix

Table S13: C-index under DeepHit and Cox proportional-hazards survival heads. Each entry reports DeepHit/CoxPH C-index at the 100% local training fraction for the same frozen TRACER representation. Heads were trained using the corresponding cohort’s training partition and evaluated on its test partition under the local head-training protocol. Values below 0.5 indicate worse-than-chance concordance under the risk-score orientation used in the analysis. D, TRACER; E, TRACER-Encoder.

Disease endpoint	In-domain	UC5	UC6
104-week horizon
Heart failure	$\mathbf{+0.048}\;(\mathbf{+0.031},\,\mathbf{+0.067})$	$\mathbf{+0.031}\;(\mathbf{+0.028},\,\mathbf{+0.035})$	$\mathbf{+0.024}\;(\mathbf{+0.020},\,\mathbf{+0.028})$
Acute myocardial infarction	$\mathbf{+0.080}\;(\mathbf{+0.050},\,\mathbf{+0.117})$	$\mathbf{+0.070}\;(\mathbf{+0.065},\,\mathbf{+0.076})$	$\mathbf{+0.051}\;(\mathbf{+0.035},\,\mathbf{+0.068})$
Lupus	$\mathbf{+0.170}\;(\mathbf{+0.071},\,\mathbf{+0.274})$	$\mathbf{+0.127}\;(\mathbf{+0.114},\,\mathbf{+0.141})$	$\mathbf{+0.070}\;(\mathbf{+0.043},\,\mathbf{+0.095})$
Hypertension	$\mathbf{+0.027}\;(\mathbf{+0.019},\,\mathbf{+0.035})$	$\mathbf{+0.008}\;(\mathbf{+0.006},\,\mathbf{+0.009})$	$\mathbf{+0.033}\;(\mathbf{+0.031},\,\mathbf{+0.036})$
Celiac disease	$+0.017\;(-0.066,\,+0.112)$	$\mathbf{+0.051}\;(\mathbf{+0.039},\,\mathbf{+0.065})$	$+0.007\;(-0.015,\,+0.027)$

Appendix

Table S16: Paired differences in time-dependent AUROC between TRACER and TRACER-Encoder at the 104-, 260-, and 26-week horizons. Differences are TRACER minus TRACER-Encoder with 95% paired patient-level bootstrap intervals; boldface indicates intervals excluding zero. The 52-week horizon is in Supplementary Table S3 .

		In-domain			UC5			UC6
Disease	Horizon (wk)	D	E	$\Delta$ (95% CI)	D	E	$\Delta$ (95% CI)	D	E	$\Delta$ (95% CI)
Chronic disease endpoints
Heart failure	52	0.692	0.672	+0.020 (+0.003, +0.037)	0.655	0.672	-0.017 (-0.020, -0.013) †	0.640	0.649	-0.009 (-0.014, -0.005) †
	104	0.686	0.659	+0.027 (+0.011, +0.043)	0.662	0.674	-0.012 (-0.015, -0.009) †	0.635	0.648	-0.013 (-0.017, -0.009) †
	260	0.677	0.649	+0.028 (+0.008, +0.048)	0.683	0.683	-0.001 (-0.004, +0.003)	0.631	0.652	-0.021 (-0.026, -0.016) †
Acute myocardial infarction	52	0.664	0.640	+0.024 (-0.017, +0.064)	0.652	0.631	+0.022 (+0.015, +0.027)	0.668	0.650	+0.018 (+0.001, +0.035)

Appendix

Table S17: Standardized partial AUROC (FPR

\leq 0.10

) for both models with paired differences. Standardized pAUC for TRACER (D) and TRACER-Encoder (E) at each prediction horizon and cohort, with the paired difference (D minus E) and its 95% paired bootstrap confidence interval. Bold: the interval excludes zero. † : strict reversal, in which the paired AUROC difference (Supplementary Tables S3 and S16 ) and the pAUC difference both exclude zero and favor different models.

	TRACER				TRACER-Encoder
Disease endpoint	52 wk	104 wk	260 wk	$\Delta_{52\rightarrow 260}$	52 wk	104 wk	260 wk	$\Delta_{52\rightarrow 260}$
Small 52-to-260-week change under TRACER ( $\lvert\Delta\rvert\leq 0.025$ )
Lupus	0.778	0.807	0.795	$+0.017$	0.694	0.637	0.675	$-0.019$
Atrial fibrillation	0.852	0.861	0.859	$+0.008$	0.774	0.793	0.778	$+0.005$
Rheumatoid arthritis	0.786	0.778	0.784	$-0.002$	0.654	0.703	0.754	$+0.100$
Heart failure	0.870	0.869	0.862	$-0.008$	0.829	0.821	0.822	$-0.007$

Appendix

Table S20: Time-dependent AUROC across prediction horizons in the in-domain test cohort. Point estimates are shown at the 52-, 104-, and 260-week horizons for survival heads trained on the full development-domain downstream training cohort. The change across the common horizon range was calculated as

\Delta_{52\rightarrow 260}=\mathrm{AUROC}_{260}-\mathrm{AUROC}_{52}

; positive values indicate higher AUROC at 260 weeks, whereas negative values indicate lower AUROC. Endpoints are grouped according to the magnitude of the 52-to-260-week change under TRACER.

		TRACER					TRACER-Encoder
Disease	Cohort	10%	30%	50%	100%	$\Delta$	10%	30%	50%	100%	$\Delta$
Heart failure	In-domain	0.837	0.858	0.862	0.869	$+0.032$	0.722	0.766	0.781	0.810	$+0.087$
	UC5	0.851	0.854	0.858	0.865	$+0.014$	0.754	0.797	0.807	0.823	$+0.070$
	UC6	0.793	0.803	0.803	0.836	$+0.043$	0.721	0.766	0.780	0.798	$+0.078$
Acute myocardial infarction	In-domain	0.788	0.837	0.830	0.842	$+0.054$	0.709	0.735	0.736	0.775	$+0.066$
	UC5	0.801	0.830	0.820	0.830	$+0.029$	0.719	0.737	0.737	0.776	$+0.057$

Appendix

Table S22: Uno’s IPCW C-index by in-domain training fraction in all three evaluation cohorts. Survival heads trained on 10%, 30%, 50%, and 100% of the in-domain downstream training cohort and applied unchanged to each cohort (original development heads).

\Delta

is the 100% minus 10% difference. The in-domain rows reproduce Supplementary Table S21 .

Figure S1: Global and low-false-positive discrimination yielded different model rankings in a subset of evaluations All paired differences were calculated as TRACER minus TRACER-Encoder; positive values favor TRACER, whereas negative values favor TRACER-Encoder. (a) Paired differences in dynamic AUROC and standardized partial AUROC over the prespecified low-false-positive region (pAUC; FPR, 0-0.1) across the 16 endpoints and the shared 52-, 104-, and 260-week horizons, shown separately for in-domain, UC5, and UC6 cohorts. Shaded quadrants indicate evaluations in which the two metrics produced opposite point-estimate model rankings. Filled points denote ranking disagreements for which the paired 95% bootstrap confidence interval for the pAUC difference excluded zero. (b) Endpoint-by-horizon map of point-estimate agreement and disagreement between AUROC and pAUC-based model rankings. Gray cells indicate agreement and orange cells indicate disagreement. Asterisks denote disagreements for which the paired 95% confidence interval for the pAUC difference excluded zero; 15 of these 16 cells also have AUROC confidence intervals excluding zero and therefore met the strict-reversal criterion. Plus signs indicate the less common reverse direction, in which pAUC favored TRACER and AUROC favored TRACER-Encoder. (c) Paired AUROC differences at the 104-week horizon for each endpoint in the in-domain, UC5, and UC6 cohorts. Points and horizontal bars show the paired difference and its 95% bootstrap confidence interval, respectively; filled points indicate intervals excluding zero.

Figure S2: Interpretability analyses reveal token-level and site-level distribution characteristics . (a) Mean integrated-gradients attribution mass for the paired difference in predicted risk between TRACER and TRACER-Encoder, aggregated by token family across six selected chronic disease endpoints. Darker shading indicates greater attribution mass. Cell values represent each token family’s share of normalized attribution mass and are not adjusted for token frequency. (b) Attribution enrichment relative to token frequency for laboratory, diagnosis, medication, and procedure tokens. Enrichment was calculated by comparing each family’s share of attribution mass with its share of input-token positions; values greater than 1 indicate that a token family received more attribution than expected from its frequency. Structural tokens are not included in the enrichment panel. (c) Site-level characteristics of the in-domain, UC5, and UC6 evaluation cohorts, including Jensen-Shannon divergence of clinical-token frequencies from the in-domain reference, median sequence length, mean number of unique tokens per patient, median follow-up in weeks, and overall event prevalence. Yellow, blue, and orange indicate the in-domain, UC5, and UC6 cohorts, respectively.

EHR Foundation Models

See all 22 papers

May 14, 2026

DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System

Accurate disease trajectory prediction is critical for early intervention, resource allocation, and improving long-term outcomes. While electronic health records (EHRs) provide a rich longitudinal view of patient health in clinical environments, models trained on curated research cohorts may not reflect routine deployment settings, and those trained on single-hospital datasets capture only fragments of each patient's trajectory. This highlights the importance of leveraging large, multi-hospital health systems for training and validation to better reflect real-world clinical complexity. In this work, we develop DT-Transformer, a foundation model trained on 57.1M structured EHR entries over 1.7M patients from Mass General Brigham (MGB), spanning 11 hospitals and a broad network of outpatient clinics. DT-Transformer achieves strong discrimination in both held-out and prospective validation settings. Next-event prediction achieves a median age- and sex-stratified AUC of 0.871 across 896 disease categories, with all categories exceeding AUC 0.5. These results support health system-scale training as a path toward foundation models suited to real-world clinical forecasting.

Yunying Zhu, Andrew R Weckstein, Kueiyu Joshua Lin +1

Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA · Harvard T.H. Chan School of Public Health, Harvard University, Boston, MA, USA · Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, MA, USA +2

Clinical Prediction Time Series Transformers Time-Series Foundation Models Healthcare Longitudinal EHR Modeling EHR Foundation Models+1

Ask PDF

Mar 9, 2026

EveryQuery: A Promptable Foundation Model for Clinical Prediction Tasks over Electronic Health Records

Autoregressive foundation models for electronic health records can make predictions for diverse clinical tasks without finetuning. These methods work by generating possible "synthetic futures" for patients, and then inferring downstream predictions over these simulated trajectories. Despite their strong capabilities, these methods have several flaws. (1) Inference is extremely computationally expensive, as each prediction requires simulating many trajectories, each with many observations. (2) Their performance is noisy due to the variance inherent to simulation, which in particular reduces efficacy on rare events, which are often of high importance clinically. (3) They are not promptable, as their only input is the patient's medical history, requiring indirect aggregation at inference time. We introduce EveryQuery, a promptable foundation model for structured EHR data. Building on the insight that clinical prediction tasks can be written in a structured query language, EveryQuery takes a query as a prompt alongside the patient's history and estimates the outcome directly. This allows EveryQuery to make predictions across diverse clinical tasks, such as classification and survival modeling, without simulation or finetuning. Across three EHR datasets, EveryQuery outperforms a competitive autoregressive baseline on clinically relevant classification and time-to-event tasks, with mean AUROC gains of 14.3% and 12.7% on MIMIC-IV and a large academic medical center and comparable performance on the smaller NWICU dataset. Its advantage is largest for rare outcomes, and it requires 580x less inference compute.

Payal Chandak, Gregory Kondas, Liat Antwarg Friedman +2

Harvard-MIT HST · Columbia University · Harvard Medical School

EHR Foundation Models Clinical Outcome Prediction Clinical Prediction Longitudinal EHR Modeling Electronic Health Records

Ask PDF

Aug 17, 2026

Foresight-England: Development of a National-Scale Generative AI Model of Electronic Health Records for Medical Event Prediction across the COVID-19 Pandemic

Foresight-England (Foresight-E) is the first national-scale generative foundation model of electronic health records (EHRs), developed as a research pilot strictly for COVID-19 research. We evaluated its ability to model the direct and indirect effects of the pandemic. Trained from scratch entirely within the NHS England Secure Data Environment, Foresight-E is a 243-million-parameter transformer decoder. It was trained and evaluated on de-identified, longitudinal EHRs of approximately 61 million individuals, integrating primary/secondary care, death registrations, and COVID-19 data. Training and validation used a 90% subset (54.9 million) spanning November 2018 to December 2022; the remaining 10% (6.1 million) was held out for evaluation. Foresight-E models patient timelines autoregressively, predicting the next medical event given their prior history. At inference, it operates zero-shot, predicting any concept in its ~40,000-code vocabulary without task-specific training. Our tokenisation scheme retains the clinical granularity of ICD-10, OPCS-4, and SNOMED CT codes, jointly representing absolute and relative timing. We designed an evaluation framework for 30-day COVID-19 hospitalisation and mortality, including subgroup analyses by demographic factors and vaccination status. To assess generalisation to unseen future data and the pandemic's indirect effects, we tested the model on medical events from 2023 (beyond its training period), benchmarking against logistic regression and XGBoost. As detailed in the Project Status section, NHS England has paused access to data for the Foresight-E project, meaning quantitative results are currently unavailable. Instead, we share our strategy for tokenisation, architecture, training, inference, and evaluation as a methodological template and case study in the challenges of building population-scale EHR foundation models.

Simon Ellershaw, Christopher Tomlinson, Zeljko Kraljevic +6

Institute of Health Informatics, University College London, London, UK · Department of Biostatistics and Health Informatics, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK · King’s Institute for Artificial Intelligence, King’s College London, London, UK +13

EHR Foundation Models Next-Event Prediction Longitudinal EHR Modeling Electronic Health Records Clinical Outcome Prediction

Ask PDF

Abstract

EHR Foundation Models

Explore similar work

DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System

EveryQuery: A Promptable Foundation Model for Clinical Prediction Tasks over Electronic Health Records

Foresight-England: Development of a National-Scale Generative AI Model of Electronic Health Records for Medical Event Prediction across the COVID-19 Pandemic