Hagai Rossman
Research Subjects
Research Interests
Hagai Rossman has a total of 149 co-authors. Here are the top co-authors by number of shared publications:
- Eran Segal
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
24 shared publications - Smadar Shilo
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
12 shared publications Anastasia Godneva
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
8 shared publicationsSmadar Shilo
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
8 shared publicationsTomer Meir
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
5 shared publicationsYeela Talmor-Barkan
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel
5 shared publications
Latest Publications
- 19 August 2026
- Published by Springer Nature in Nature Health
- p. 1-12
Impact of daily diet on sleep quality
- 19 August 2026
- Published by Springer Nature in Nature Health
- p. 1-12
Impact of daily diet on sleep quality
- Open Access
Grounding Health AI: Architecture and Evaluation of a Domain-Expert Metabolic Health Agent
General-purpose language models generate fluent health reports that can fabricate derived clinical met-rics. In an illustrative comparison on identical two-week CGM and meal data, leading foundation models produced reports with invented MAGE values, inflated meal counts, and unreferenced complication-risk projections — failures invisible to non-expert readers and plausible enough to mislead clinicians. We describe the HPP Personal Health Agent (PHA), a metabolic health agent that grounds generation in four layers: the Human Phenotype Project (HPP), a deep-phenotyped cohort of 13,000+ participants sup-plying population references and trained predictive models; 21 domain-expert tools and trained-model wrappers that compute clinical metrics and risk predictions; declarative behavioural skills that constrain what the model may claim; and 21 automated evals across 8 categories developed via a test-driven cycle in which each eval encodes a failure mode discovered during iterative development. In a 210-report ma-trix (14 participants × 3 prompts × 5 system conditions), the gains are largest on the system’s primary use case — meal-grounded metabolic reports, the report it was designed for — where the full system raises a deterministic form/provenance score from 0.37 (the same foundation model with no tools or skills) to 0.91; this score measures structural completeness, numerical accuracy, tool grounding, and clinical-language compliance — a necessary condition for trustworthy health reporting, with clinical quality as a complementary axis examined qualitatively. A skills-vs-tools decomposition shows the two layers act on different axes: tools drive numerical accuracy (≈14% → 90% of reported metrics correct), while the declarative skills add most of the remaining gain in citations, completeness, and structure (tools alone recover only part of the gap, 0.49 from the same 0.37 baseline). The lift generalises beyond the primary use case — to a second metabolic prompt (0.72) and a cardiovascular extension (0.70), each from a 0.37–0.39 baseline. The architecture extends across clinical domains: adding a SCORE2 cardiovascular risk tool and a corresponding skill — with no changes to orchestration, eval harness, or existing tools — produced a cardiovascular risk report from the same system. Trustworthy domain-specialised health AI is a systems design problem: deep-phenotyped cohort data, domain-expert tools and models, and eval-driven development together form a replicable pattern.
Open AccessGrounding Health AI: Architecture and Evaluation of a Domain-Expert Metabolic Health Agent
General-purpose language models generate fluent health reports that can fabricate derived clinical met-rics. In an illustrative comparison on identical two-week CGM and meal data, leading foundation models produced reports with invented MAGE values, inflated meal counts, and unreferenced complication-risk projections — failures invisible to non-expert readers and plausible enough to mislead clinicians. We describe the HPP Personal Health Agent (PHA), a metabolic health agent that grounds generation in four layers: the Human Phenotype Project (HPP), a deep-phenotyped cohort of 13,000+ participants sup-plying population references and trained predictive models; 21 domain-expert tools and trained-model wrappers that compute clinical metrics and risk predictions; declarative behavioural skills that constrain what the model may claim; and 21 automated evals across 8 categories developed via a test-driven cycle in which each eval encodes a failure mode discovered during iterative development. In a 210-report ma-trix (14 participants × 3 prompts × 5 system conditions), the gains are largest on the system’s primary use case — meal-grounded metabolic reports, the report it was designed for — where the full system raises a deterministic form/provenance score from 0.37 (the same foundation model with no tools or skills) to 0.91; this score measures structural completeness, numerical accuracy, tool grounding, and clinical-language compliance — a necessary condition for trustworthy health reporting, with clinical quality as a complementary axis examined qualitatively. A skills-vs-tools decomposition shows the two layers act on different axes: tools drive numerical accuracy (≈14% → 90% of reported metrics correct), while the declarative skills add most of the remaining gain in citations, completeness, and structure (tools alone recover only part of the gap, 0.49 from the same 0.37 baseline). The lift generalises beyond the primary use case — to a second metabolic prompt (0.72) and a cardiovascular extension (0.70), each from a 0.37–0.39 baseline. The architecture extends across clinical domains: adding a SCORE2 cardiovascular risk tool and a corresponding skill — with no changes to orchestration, eval harness, or existing tools — produced a cardiovascular risk report from the same system. Trustworthy domain-specialised health AI is a systems design problem: deep-phenotyped cohort data, domain-expert tools and models, and eval-driven development together form a replicable pattern.
- Open Access
A foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk
While Photoplethysmography (PPG) is established as a noninvasive optical tool for monitoring heart rate and oxygen saturation, its high-resolution pulsatile blood volume waveforms contain rich physiological data that extend far beyond conventional vital signs. We introduce PulseOx-FM, a foundation model, trained using self-supervised learning on 6,995,558 segments of pulse oximetry signals collected during 42,282 overnight sleep monitoring recordings of 10,704 adult participants in the Human Phenotype Project (HPP). Using chronological age as a global health benchmark, PulseOx-FM significantly outperformed existing open-source and proprietary feature extraction methods while demonstrating robust generalization in an external out-of-distribution cohort. PulseOx-FM representations predicted 71 phenotypic targets spanning cardiometabolic, and neuropsychiatric domains beyond age, sex and body mass index. Applied in the same way to 3,721 intensive-care patients, two-minute windows of routine bedside pulse oximetry from the first day of admission improved the prediction of survival beyond age, sex and the heart rate and oxygen saturation the same device reports. Nightly embeddings further tracked next-day glycemic, dietary and activity-based state within individuals, dissociating this signal from sleep architecture alone. This next-day glycemic signal was predominantly a direct physiological effect, not explained by next-day dietary intake. These findings suggest that PulseOx-FM provides a generalizable framework for encoding physiological patterns from sleep, offering a non-invasive tool for global health risk stratification and precision medicine.
Open AccessA foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk
While Photoplethysmography (PPG) is established as a noninvasive optical tool for monitoring heart rate and oxygen saturation, its high-resolution pulsatile blood volume waveforms contain rich physiological data that extend far beyond conventional vital signs. We introduce PulseOx-FM, a foundation model, trained using self-supervised learning on 6,995,558 segments of pulse oximetry signals collected during 42,282 overnight sleep monitoring recordings of 10,704 adult participants in the Human Phenotype Project (HPP). Using chronological age as a global health benchmark, PulseOx-FM significantly outperformed existing open-source and proprietary feature extraction methods while demonstrating robust generalization in an external out-of-distribution cohort. PulseOx-FM representations predicted 71 phenotypic targets spanning cardiometabolic, and neuropsychiatric domains beyond age, sex and body mass index. Applied in the same way to 3,721 intensive-care patients, two-minute windows of routine bedside pulse oximetry from the first day of admission improved the prediction of survival beyond age, sex and the heart rate and oxygen saturation the same device reports. Nightly embeddings further tracked next-day glycemic, dietary and activity-based state within individuals, dissociating this signal from sleep architecture alone. This next-day glycemic signal was predominantly a direct physiological effect, not explained by next-day dietary intake. These findings suggest that PulseOx-FM provides a generalizable framework for encoding physiological patterns from sleep, offering a non-invasive tool for global health risk stratification and precision medicine.
- Open Access
Pre-activity glycemic prediction prioritizes post-meal movement
Post-meal activity can attenuate glucose excursions; however, the exact magnitude of this effect remains unquantified, and guidance is rarely personalized to the meal occasion. We linked Human Phenotype Project diet logs, continuous glucose monitoring and wearable step counts to test whether glycemic risk estimated before activity occurs can prioritize post-meal movement. An activity-blind PPGR model trained on 391,214 PPGR-valid meals from 9,561 participants generated pre-activity meal scores. Among 55,949 step-linked meals from 1,627 adults without diabetes, higher 0-120-min post-meal steps were associated with lower within-participant PPGR (-53.0 mg/dL*min per 1 s.d. higher log steps; 95% CI, -64.2 to -41.7), with larger adjusted PPGR iAUC contrasts at 1,501-2,500 observed steps (-154.4 mg/dL*min versus 0-50 steps). Associations were stronger among participants with higher glycemic-adiposity burden and after meals with higher predicted PPGR. A held-out pre-activity step-response ranking concentrated larger inverse step-PPGR associations (-79.1 top versus -15.0 mg/dL*min bottom quintile), providing a testable strategy for prediction-guided, post-meal movement prompts.
Open AccessPre-activity glycemic prediction prioritizes post-meal movement
Post-meal activity can attenuate glucose excursions; however, the exact magnitude of this effect remains unquantified, and guidance is rarely personalized to the meal occasion. We linked Human Phenotype Project diet logs, continuous glucose monitoring and wearable step counts to test whether glycemic risk estimated before activity occurs can prioritize post-meal movement. An activity-blind PPGR model trained on 391,214 PPGR-valid meals from 9,561 participants generated pre-activity meal scores. Among 55,949 step-linked meals from 1,627 adults without diabetes, higher 0-120-min post-meal steps were associated with lower within-participant PPGR (-53.0 mg/dL*min per 1 s.d. higher log steps; 95% CI, -64.2 to -41.7), with larger adjusted PPGR iAUC contrasts at 1,501-2,500 observed steps (-154.4 mg/dL*min versus 0-50 steps). Associations were stronger among participants with higher glycemic-adiposity burden and after meals with higher predicted PPGR. A held-out pre-activity step-response ranking concentrated larger inverse step-PPGR associations (-79.1 top versus -15.0 mg/dL*min bottom quintile), providing a testable strategy for prediction-guided, post-meal movement prompts.
- Open Access
- 18 June 2026
- Published by Elsevier in JACC: Basic To Translational Science
- Vol. 11 (7) , 101596
RetiMap
- 18 June 2026
- Published by Elsevier in JACC: Basic To Translational Science
- Vol. 11 (7) , 101596
Open AccessRetiMap
- Top Cited Papers
- 23 March 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (5) , 1884-1894
Diet–microbiome associations in 10,068 individuals from the Human Phenotype Project to guide personalized nutrition
Diet is a major environmental factor influencing the human gut microbiome. However, the effects of specific foods and dietary patterns on microbial composition, diversity and function is not fully understood, limiting progress toward personalized dietary strategies. Here, leveraging 10,068 participants from the Human Phenotype Project with app-based diet logs and shotgun metagenomics, we predicted diet–microbiome associations at species-level resolution. Diet significantly predicted microbial diversity (richness r = 0.26, Shannon Index r = 0.24), the relative abundance of 669 of 724 species tested (92.4%, false discovery rate Lawsonibacter asaccharolyticus (r = 0.43), yogurt with Streptococcus thermophilus (r = 0.42) and milk with Bifidobacterium species (r = 0.31–0.36). In parallel, broader dietary patterns, especially the degree of food processing, emerged as predictors of microbial diversity and composition. We also show that diet–microbiome associations persist over four years, with 82.5% of species exhibiting significant longitudinal tracking between predicted and observed abundances. Finally, we developed an exploratory analysis for simulating personalized dietary interventions with predicted microbiome shift effects that are associated with improvements in cardiometabolic health. Our findings demonstrate that diet is strongly associated with microbiome composition, diversity and function, and highlight its potential for guiding personalized interventions.
- 23 March 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (5) , 1884-1894
Top Cited PapersDiet–microbiome associations in 10,068 individuals from the Human Phenotype Project to guide personalized nutrition
Diet is a major environmental factor influencing the human gut microbiome. However, the effects of specific foods and dietary patterns on microbial composition, diversity and function is not fully understood, limiting progress toward personalized dietary strategies. Here, leveraging 10,068 participants from the Human Phenotype Project with app-based diet logs and shotgun metagenomics, we predicted diet–microbiome associations at species-level resolution. Diet significantly predicted microbial diversity (richness r = 0.26, Shannon Index r = 0.24), the relative abundance of 669 of 724 species tested (92.4%, false discovery rate Lawsonibacter asaccharolyticus (r = 0.43), yogurt with Streptococcus thermophilus (r = 0.42) and milk with Bifidobacterium species (r = 0.31–0.36). In parallel, broader dietary patterns, especially the degree of food processing, emerged as predictors of microbial diversity and composition. We also show that diet–microbiome associations persist over four years, with 82.5% of species exhibiting significant longitudinal tracking between predicted and observed abundances. Finally, we developed an exploratory analysis for simulating personalized dietary interventions with predicted microbiome shift effects that are associated with improvements in cardiometabolic health. Our findings demonstrate that diet is strongly associated with microbiome composition, diversity and function, and highlight its potential for guiding personalized interventions.
- Open Access
Day-to-day dietary variation shapes overnight sleep physiology: a target-trial emulation in 4.8 thousand person-nights
Sleep architecture is essential for metabolic and cardiovascular health, yet the impact of day-to-day dietary variation on objective sleep physiology remains unclear. Using 4.8 thousand person-nights with real-time dietary logs and multi-stage wearable sleep recordings, we examined how prior-day nutrition relates to next-night sleep under free-living conditions. Higher fiber density was associated with increased restorative sleep, including +0.59 pp deep sleep, +0.76 pp REM sleep, −1.35 pp light sleep, and −1.14 bpm lower mean nocturnal heart rate. Greater plant diversity and higher whole-plant food intake were similarly associated with lower nocturnal heart rate (−0.72 to −0.94 bpm). Meal-timing behaviors primarily influenced sleep duration, sleep-onset latency, and autonomic tone: heavier evening meals were associated with +7.7 min longer total sleep time and +0.73 bpm higher nocturnal heart rate. In contrast, short-term variation in macronutrient energy distribution and micronutrient consumption showed no robust associations with sleep outcomes. When analyses were restricted to more extreme dietary contrasts, effect magnitudes increased while remaining directionally consistent. These findings indicate that routine daily dietary choices, particularly plant-forward composition and meal timing, have immediate and measurable effects on objective sleep architecture.
Open AccessDay-to-day dietary variation shapes overnight sleep physiology: a target-trial emulation in 4.8 thousand person-nights
Sleep architecture is essential for metabolic and cardiovascular health, yet the impact of day-to-day dietary variation on objective sleep physiology remains unclear. Using 4.8 thousand person-nights with real-time dietary logs and multi-stage wearable sleep recordings, we examined how prior-day nutrition relates to next-night sleep under free-living conditions. Higher fiber density was associated with increased restorative sleep, including +0.59 pp deep sleep, +0.76 pp REM sleep, −1.35 pp light sleep, and −1.14 bpm lower mean nocturnal heart rate. Greater plant diversity and higher whole-plant food intake were similarly associated with lower nocturnal heart rate (−0.72 to −0.94 bpm). Meal-timing behaviors primarily influenced sleep duration, sleep-onset latency, and autonomic tone: heavier evening meals were associated with +7.7 min longer total sleep time and +0.73 bpm higher nocturnal heart rate. In contrast, short-term variation in macronutrient energy distribution and micronutrient consumption showed no robust associations with sleep outcomes. When analyses were restricted to more extreme dietary contrasts, effect magnitudes increased while remaining directionally consistent. These findings indicate that routine daily dietary choices, particularly plant-forward composition and meal timing, have immediate and measurable effects on objective sleep architecture.
- Top Cited Papers
- 14 January 2026
- Published by Springer Nature in Nature
- Vol. 650 (8103) , 978-986
A foundation model for continuous glucose monitoring data
Continuous glucose monitoring (CGM) generates detailed temporal profiles of glucose dynamics, but its full potential for achieving glucose homeostasis and predicting long-term outcomes remains underutilized. Here we present GluFormer, a generative foundation model for CGM data trained with self-supervised learning on more than 10 million glucose measurements from 10,812 adults mainly without diabetes1,2. Using autoregressive prediction, the model learned representations that transferred across 19 external cohorts (n = 6,044) spanning 5 countries, 8 CGM devices and diverse pathophysiological states, including prediabetes, type 1 and type 2 diabetes, gestational diabetes and obesity. These representations provided consistent improvements over baseline blood glucose and HbA1c levels and other CGM-derived measures for forecasting glycaemic parameters3,4. In individuals with prediabetes, GluFormer stratified those likely to experience clinically significant increases in HbA1c over a 2-year period, outperforming baseline HbA1c and common CGM metrics. In a cohort of 580 adults with short-term CGM and a median follow-up of 11 years5, GluFormer identified individuals at elevated risk of diabetes and cardiovascular mortality more effectively than HbA1c. Specifically, 66% of incident diabetes cases and 69% of cardiovascular deaths occurred in the top risk quartile, compared with 7% and 0%, respectively, in the bottom quartile. In clinical trials, baseline CGM representations improved outcome prediction. A multimodal extension of the model that integrates dietary data generated plausible glucose trajectories and predicted individual glycaemic responses to food. Together, these findings indicate that GluFormer provides a generalizable framework for encoding glycaemic patterns and may inform precision medicine approaches for metabolic health.
- 14 January 2026
- Published by Springer Nature in Nature
- Vol. 650 (8103) , 978-986
Top Cited PapersA foundation model for continuous glucose monitoring data
Continuous glucose monitoring (CGM) generates detailed temporal profiles of glucose dynamics, but its full potential for achieving glucose homeostasis and predicting long-term outcomes remains underutilized. Here we present GluFormer, a generative foundation model for CGM data trained with self-supervised learning on more than 10 million glucose measurements from 10,812 adults mainly without diabetes1,2. Using autoregressive prediction, the model learned representations that transferred across 19 external cohorts (n = 6,044) spanning 5 countries, 8 CGM devices and diverse pathophysiological states, including prediabetes, type 1 and type 2 diabetes, gestational diabetes and obesity. These representations provided consistent improvements over baseline blood glucose and HbA1c levels and other CGM-derived measures for forecasting glycaemic parameters3,4. In individuals with prediabetes, GluFormer stratified those likely to experience clinically significant increases in HbA1c over a 2-year period, outperforming baseline HbA1c and common CGM metrics. In a cohort of 580 adults with short-term CGM and a median follow-up of 11 years5, GluFormer identified individuals at elevated risk of diabetes and cardiovascular mortality more effectively than HbA1c. Specifically, 66% of incident diabetes cases and 69% of cardiovascular deaths occurred in the top risk quartile, compared with 7% and 0%, respectively, in the bottom quartile. In clinical trials, baseline CGM representations improved outcome prediction. A multimodal extension of the model that integrates dietary data generated plausible glucose trajectories and predicted individual glycaemic responses to food. Together, these findings indicate that GluFormer provides a generalizable framework for encoding glycaemic patterns and may inform precision medicine approaches for metabolic health.
- Open Access
Unveiling Associations Between Retinal Microvascular Architecture and Phenotypes Across Thousands of Healthy Subjects
Fundus imaging has emerged as a powerful tool, allowing high-resolution and non-invasive visualization of the fundus microvasculature. Recent advancements in artificial intelligence (AI) have enabled the quantification of fundus microvascular characteristics, paving the way for the development of predictive markers for systemic conditions. In this study, we characterize the fundus microvasculature among 8467 healthy individuals, aged 40-70 years. We employ a fully automated segmentation tool to calculate 12 distinct vessel measures, including tortuosity density, vessel density, average width, fractal dimension, distance tortuosity, and curvature tortuosity for both arteries and veins. We provide reference values for these measures by age and sex, and explore the associations between these fundus vessel measures and various clinical features. Our analysis reveals numerous significant correlations, including connections between fundus arterial measures and cardiovascular measurements, lipid profile, sleep apnea indices, glycemic profile, body composition, and smoking status. This comprehensive analysis of fundus microvasculature anatomy in a large healthy population contributes to the growing field of oculomics and highlights the potential of fundus imaging for earlier detection and monitoring of systemic diseases such as metabolic syndrome, cardiovascular diseases and sleep disorders.
Open AccessUnveiling Associations Between Retinal Microvascular Architecture and Phenotypes Across Thousands of Healthy Subjects
Fundus imaging has emerged as a powerful tool, allowing high-resolution and non-invasive visualization of the fundus microvasculature. Recent advancements in artificial intelligence (AI) have enabled the quantification of fundus microvascular characteristics, paving the way for the development of predictive markers for systemic conditions. In this study, we characterize the fundus microvasculature among 8467 healthy individuals, aged 40-70 years. We employ a fully automated segmentation tool to calculate 12 distinct vessel measures, including tortuosity density, vessel density, average width, fractal dimension, distance tortuosity, and curvature tortuosity for both arteries and veins. We provide reference values for these measures by age and sex, and explore the associations between these fundus vessel measures and various clinical features. Our analysis reveals numerous significant correlations, including connections between fundus arterial measures and cardiovascular measurements, lipid profile, sleep apnea indices, glycemic profile, body composition, and smoking status. This comprehensive analysis of fundus microvasculature anatomy in a large healthy population contributes to the growing field of oculomics and highlights the potential of fundus imaging for earlier detection and monitoring of systemic diseases such as metabolic syndrome, cardiovascular diseases and sleep disorders.
- Open Access
COMPRER: A Multimodal Multi-Objective Pretraining Framework for Enhanced Medical Image Representation
Substantial advances in multi-modal Artificial Intelligence (AI) facilitate the combination of diverse medical modalities to achieve holistic health assessments. We present COMPRER, a novel multi-modal, multi-objective pretraining framework which enhances medical-image representation, diagnostic inferences, and prognosis of diseases. COMPRER employs a multi-objective training framework, where each objective introduces distinct knowledge to the model. This includes a multi-modal loss that consolidates information across different imaging modalities; A temporal loss that imparts the ability to discern patterns over time; Medical-measure prediction adds appropriate medical insights; Lastly, reconstruction loss ensures the integrity of image structure within the latent space. Despite the concern that multiple objectives could weaken task performance, our findings show that this combination actually boosts outcomes on certain tasks. Here, we apply this framework to both fundus images and carotid ultrasound, and validate our downstream tasks capabilities by predicting both current and future cardiovascular conditions. COMPRER achieved higher Area Under the Curve (AUC) scores in evaluating medical conditions compared to existing models on held-out data. On the Out-of-distribution (OOD) UK-Biobank dataset COMPRER maintains favorable performance over well-established models with more parameters, even though these models were trained on 75× more data than COMPRER. In addition, to better assess our model’s performance in contrastive learning, we introduce a novel evaluation metric, providing deeper understanding of the effectiveness of the latent space pairing.
Open AccessCOMPRER: A Multimodal Multi-Objective Pretraining Framework for Enhanced Medical Image Representation
Substantial advances in multi-modal Artificial Intelligence (AI) facilitate the combination of diverse medical modalities to achieve holistic health assessments. We present COMPRER, a novel multi-modal, multi-objective pretraining framework which enhances medical-image representation, diagnostic inferences, and prognosis of diseases. COMPRER employs a multi-objective training framework, where each objective introduces distinct knowledge to the model. This includes a multi-modal loss that consolidates information across different imaging modalities; A temporal loss that imparts the ability to discern patterns over time; Medical-measure prediction adds appropriate medical insights; Lastly, reconstruction loss ensures the integrity of image structure within the latent space. Despite the concern that multiple objectives could weaken task performance, our findings show that this combination actually boosts outcomes on certain tasks. Here, we apply this framework to both fundus images and carotid ultrasound, and validate our downstream tasks capabilities by predicting both current and future cardiovascular conditions. COMPRER achieved higher Area Under the Curve (AUC) scores in evaluating medical conditions compared to existing models on held-out data. On the Out-of-distribution (OOD) UK-Biobank dataset COMPRER maintains favorable performance over well-established models with more parameters, even though these models were trained on 75× more data than COMPRER. In addition, to better assess our model’s performance in contrastive learning, we introduce a novel evaluation metric, providing deeper understanding of the effectiveness of the latent space pairing.