David H. Rehkopf
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
David H. Rehkopf has a total of 373 co-authors. Here are the top co-authors by number of shared publications:
- Belinda L. Needham
Department of Epidemiology, University of Michigan School of Public Health , 1415 Washington Heights, Ann Arbor, MI 48109-2029
8 shared publications - Luis Rosero-Bixby
Centro Centroamericano de Población, Universidad de Costa Rica
8 shared publications - Mark R. Cullen
Stanford University
8 shared publications - Nicole Gladish
Stanford University
8 shared publications - Alison Klebanoff Cohen
Department of Epidemiology & Biostatistics, University of California San Francisco, San Francisco, USA
7 shared publications - Barbara Abrams
Division of Epidemiology, School of Public Health, University of California, Berkeley, CA
7 shared publications
Latest Publications
- 3 November 2026
- Published by American Association for Cancer Research (AACR) in Cancer Epidemiology, Biomarkers & Prevention
- p. OF1-OF12
Hispanic and Asian American/Pacific Islander Ethnic Enclave, Neighborhood Socioeconomic Status, and Renal Cell Carcinoma Incidence in California
Background: We evaluated associations of neighborhood socioeconomic status (nSES) and ethnic enclave with renal cell carcinoma (RCC) incidence among Hispanic and Asian American/Pacific Islander (AAPI) populations in California. Methods: We conducted a population-based study of 7,507 Hispanic and 2,293 AAPI individuals diagnosed with RCC in the California Cancer Registry across three Census periods (1988–1992, 1998–2002, and 2008–2012). Race- and ethnicity-specific Census population counts were used to estimate incidence rate ratios (IRR) and 95% confidence intervals (CI) across nSES and ethnic enclave quintiles (Q) and a joint variable of dichotomous nSES (high vs. low) and ethnic enclave (enclave vs. nonenclave). Results: Higher nSES was associated with higher RCC incidence among AAPI individuals (Q5 vs. Q1: IRR, 1.50; 95% CI, 1.27–1.78) and modestly elevated incidence among Hispanic individuals in intermediate quintiles. Less ethnically distinct neighborhoods were associated with higher incidence among Hispanic individuals (Q1 vs. Q5: IRR, 1.21; 95% CI, 1.10–1.34), whereas no clear pattern was observed among AAPI individuals. Compared with low-nSES/enclave neighborhoods, RCC incidence was higher in all other joint nSES/enclave categories in both populations. These associations were generally similar by sex, age at diagnosis, and tumor histology and were mainly observed for localized-stage RCC. Conclusions: Hispanic and AAPI individuals living in higher-nSES neighborhoods had higher RCC incidence, whereas Hispanic individuals living in less ethnically distinct neighborhoods had higher RCC incidence. These patterns may reflect differences in healthcare access, diagnostic intensity, and neighborhood-related etiologic factors. Impact: Neighborhood factors should be considered in RCC etiology, surveillance, and risk stratification research among diverse populations.
- 3 November 2026
- Published by American Association for Cancer Research (AACR) in Cancer Epidemiology, Biomarkers & Prevention
- p. OF1-OF12
Hispanic and Asian American/Pacific Islander Ethnic Enclave, Neighborhood Socioeconomic Status, and Renal Cell Carcinoma Incidence in California
Background: We evaluated associations of neighborhood socioeconomic status (nSES) and ethnic enclave with renal cell carcinoma (RCC) incidence among Hispanic and Asian American/Pacific Islander (AAPI) populations in California. Methods: We conducted a population-based study of 7,507 Hispanic and 2,293 AAPI individuals diagnosed with RCC in the California Cancer Registry across three Census periods (1988–1992, 1998–2002, and 2008–2012). Race- and ethnicity-specific Census population counts were used to estimate incidence rate ratios (IRR) and 95% confidence intervals (CI) across nSES and ethnic enclave quintiles (Q) and a joint variable of dichotomous nSES (high vs. low) and ethnic enclave (enclave vs. nonenclave). Results: Higher nSES was associated with higher RCC incidence among AAPI individuals (Q5 vs. Q1: IRR, 1.50; 95% CI, 1.27–1.78) and modestly elevated incidence among Hispanic individuals in intermediate quintiles. Less ethnically distinct neighborhoods were associated with higher incidence among Hispanic individuals (Q1 vs. Q5: IRR, 1.21; 95% CI, 1.10–1.34), whereas no clear pattern was observed among AAPI individuals. Compared with low-nSES/enclave neighborhoods, RCC incidence was higher in all other joint nSES/enclave categories in both populations. These associations were generally similar by sex, age at diagnosis, and tumor histology and were mainly observed for localized-stage RCC. Conclusions: Hispanic and AAPI individuals living in higher-nSES neighborhoods had higher RCC incidence, whereas Hispanic individuals living in less ethnically distinct neighborhoods had higher RCC incidence. These patterns may reflect differences in healthcare access, diagnostic intensity, and neighborhood-related etiologic factors. Impact: Neighborhood factors should be considered in RCC etiology, surveillance, and risk stratification research among diverse populations.
- 15 September 2026
- Published by Elsevier in Annals of Epidemiology
- Vol. 122
Sexual orientation trajectory clusters and depression risk in emerging adulthood
- 15 September 2026
- Published by Elsevier in Annals of Epidemiology
- Vol. 122
Sexual orientation trajectory clusters and depression risk in emerging adulthood
- Open Access
- 1 July 2026
- Published by Oxford University Press (OUP) in Research Connections
- Vol. 1 (3)
Retrospective cohort study measuring the disparities in post-COVID conditions among persons with disabilities: findings from primary care settings, 2019–22
Background and aims: This study compared the incidence of post-COVID conditions by pre-pandemic disability status within primary care. Methods: This retrospective observational study used electronic health record data to identify primary care patients with ≥1 visit during January 2018–December 2019 and ≥1 visit during May 2020–December 2022. Disability status was identified using diagnostic codes from January 2018 to December 2019. COVID-19 diagnosis was captured during April 2020–December 2022. Accelerated Failure Time models were used to assess time to incident circulatory, endocrine, neurological, and respiratory diagnoses at least 28 days post COVID-19. Results: Among patients aged ≥50 years, those with disabilities and COVID-19 had higher incidence and earlier onset of newly diagnosed circulatory [time ratio (TR) 0.70, 95% confidence interval (CI) 0.64–0.77], endocrine (TR 0.91, 95% CI 0.84–0.99), and neurological (TR 0.52, 95% CI 0.44–0.62), but not respiratory conditions during the post-COVID period as compared to patients with COVID-19 without disability. No associations were found among patients aged <50 years. Conclusion: Primary care patients with pre-existing disability experienced a higher incidence and earlier onset of post-COVID conditions than patients without disability.
- 1 July 2026
- Published by Oxford University Press (OUP) in Research Connections
- Vol. 1 (3)
Open AccessRetrospective cohort study measuring the disparities in post-COVID conditions among persons with disabilities: findings from primary care settings, 2019–22
Background and aims: This study compared the incidence of post-COVID conditions by pre-pandemic disability status within primary care. Methods: This retrospective observational study used electronic health record data to identify primary care patients with ≥1 visit during January 2018–December 2019 and ≥1 visit during May 2020–December 2022. Disability status was identified using diagnostic codes from January 2018 to December 2019. COVID-19 diagnosis was captured during April 2020–December 2022. Accelerated Failure Time models were used to assess time to incident circulatory, endocrine, neurological, and respiratory diagnoses at least 28 days post COVID-19. Results: Among patients aged ≥50 years, those with disabilities and COVID-19 had higher incidence and earlier onset of newly diagnosed circulatory [time ratio (TR) 0.70, 95% confidence interval (CI) 0.64–0.77], endocrine (TR 0.91, 95% CI 0.84–0.99), and neurological (TR 0.52, 95% CI 0.44–0.62), but not respiratory conditions during the post-COVID period as compared to patients with COVID-19 without disability. No associations were found among patients aged <50 years. Conclusion: Primary care patients with pre-existing disability experienced a higher incidence and earlier onset of post-COVID conditions than patients without disability.
- Open Access
Nickel and Dimed: How a Common Earth Element is Short-Changing Our Health
Nickel has been studied for a long time as an environmental contaminant but less so in its connection to population health. It does not announce itself as loudly as its transition metal brethren like mercury and cadmium, but its chemical properties permit it to be deleterious as a low-dose, chronic exposure, particularly among those with immune systems sensitized to it. There is a growing evidence base and vocabulary to discuss nickel’s effect on health. However, in the U.S., there are not recent, reliable estimates of the share of the population with a nickel allergy, let alone how much nickel Americans are exposed to through their diet. This paper seeks to close this evidence gap by creating a new dataset of dietary nickel and other heavy metal exposure and assessing how high levels of dietary nickel exposure shape local demand for health care services. We use soil data from the U.S. Geological Survey and data on agricultural product transport from FoodFlows.org to create a county-level dietary nickel exposure index. We then use a large electronic health record database and double machine learning to estimate how demand for primary care services varies across levels of dietary nickel exposure. We find that counties with high nickel exposure experience an increase in the share of primary care office visits for symptoms highly suggestive of nickel poisoning. This result survives multiple hypothesis test corrections and placebo tests. Our research suggests that nickel has harmful effects on individual health whose exposure can be measured at a population level, and is shaping primary care across the U.S.
Open AccessNickel and Dimed: How a Common Earth Element is Short-Changing Our Health
Nickel has been studied for a long time as an environmental contaminant but less so in its connection to population health. It does not announce itself as loudly as its transition metal brethren like mercury and cadmium, but its chemical properties permit it to be deleterious as a low-dose, chronic exposure, particularly among those with immune systems sensitized to it. There is a growing evidence base and vocabulary to discuss nickel’s effect on health. However, in the U.S., there are not recent, reliable estimates of the share of the population with a nickel allergy, let alone how much nickel Americans are exposed to through their diet. This paper seeks to close this evidence gap by creating a new dataset of dietary nickel and other heavy metal exposure and assessing how high levels of dietary nickel exposure shape local demand for health care services. We use soil data from the U.S. Geological Survey and data on agricultural product transport from FoodFlows.org to create a county-level dietary nickel exposure index. We then use a large electronic health record database and double machine learning to estimate how demand for primary care services varies across levels of dietary nickel exposure. We find that counties with high nickel exposure experience an increase in the share of primary care office visits for symptoms highly suggestive of nickel poisoning. This result survives multiple hypothesis test corrections and placebo tests. Our research suggests that nickel has harmful effects on individual health whose exposure can be measured at a population level, and is shaping primary care across the U.S.
- Open Access
Diabetes and the Life-Course: Evidence from Panel Data and Electronic Health Records
Diabetes screening can alter both clinical care and patient behavior, but its consequences for health and socioeconomic outcomes remain uncertain. We examine these consequences in a fuzzy regression discontinuity design using electronic health records from the American Family Cohort (AFC) and longitudinal data from the Health and Retirement Survey (HRS). In AFC, we find that the diagnosis discontinuity peaks at 13 percentage points within six months and attenuates until it becomes insignificant at 42 months. We find a smaller and shorter-lived discontinuity for related prescriptions. These discontinuities remain across alternative specifications. Apart from specification and treatment window-dependent improvements in weight, BMI, and serum glucose, we do not observe a pattern of sustained improvement in cardiometabolic outcomes as a result of diagnosis and related care. In HRS, the threshold induces reductions of 22 percentage points in working for pay and 20.6 percentage points in labor-force participation at the two year follow-up. We do not observe a corresponding change in retirement, suggesting that diabetes diagnoses may lead to respondents reducing work without retirement. These reductions remain after several sensitivity checks but do not persist beyond two years. These findings suggest that health screening for chronic conditions may temporarily reduce labor supply, making employment responses relevant to the evaluation of broader costs and benefits of screening.
Open AccessDiabetes and the Life-Course: Evidence from Panel Data and Electronic Health Records
Diabetes screening can alter both clinical care and patient behavior, but its consequences for health and socioeconomic outcomes remain uncertain. We examine these consequences in a fuzzy regression discontinuity design using electronic health records from the American Family Cohort (AFC) and longitudinal data from the Health and Retirement Survey (HRS). In AFC, we find that the diagnosis discontinuity peaks at 13 percentage points within six months and attenuates until it becomes insignificant at 42 months. We find a smaller and shorter-lived discontinuity for related prescriptions. These discontinuities remain across alternative specifications. Apart from specification and treatment window-dependent improvements in weight, BMI, and serum glucose, we do not observe a pattern of sustained improvement in cardiometabolic outcomes as a result of diagnosis and related care. In HRS, the threshold induces reductions of 22 percentage points in working for pay and 20.6 percentage points in labor-force participation at the two year follow-up. We do not observe a corresponding change in retirement, suggesting that diabetes diagnoses may lead to respondents reducing work without retirement. These reductions remain after several sensitivity checks but do not persist beyond two years. These findings suggest that health screening for chronic conditions may temporarily reduce labor supply, making employment responses relevant to the evaluation of broader costs and benefits of screening.
- Open Access
- 1 April 2026
- Published by Oxford University Press (OUP) in AJE Advances: Research in Epidemiology
- Vol. 2 (3)
Place-based disorder, social cohesion, and physical and mental health outcomes of LGBTQIA+ adults in the United States
Place-based disorder and social cohesion may influence lesbian, gay, bisexual, transgender, queer, intersex, asexual, and other sexual and gender minority (LGBTQIA+) adults’ physical and mental health in meaningful ways given their heightened exposure to minority stress and discrimination. However, few studies have examined these associations. Using a sample of 3790 LGBTQIA+ adults in 786 counties from The Population Research in Identity and Disparities for Equality Study, we assessed associations of place-based disorder and social cohesion with depressive symptoms, perceived stress, and physical health using linear mixed models. We tested effect modification by gender modality and ethnoracial group. Residing in counties with higher social cohesion by 1 unit was associated with 1.06 points lower (95% CI, −1.56, −0.56) for depressive symptoms, 1.60 points lower (95% CI, −2.26, −0.94) for perceived stress, and 1.17 points higher (95% CI, 0.43, 1.90) for physical health. Residing in counties with higher place-based disorder by 1 unit was associated with 1.17 points higher (95% CI, 0.32, 2.01) for perceived stress; no association was observed for depressive symptoms or physical health. We did not detect evidence of effect modification by ethnoracial groups, though the association between place-based disorder and depressive symptoms appeared to differ by gender modality. Findings indicate that physical and social environments are important to the health of LGBTQIA+ individuals.
- 1 April 2026
- Published by Oxford University Press (OUP) in AJE Advances: Research in Epidemiology
- Vol. 2 (3)
Open AccessPlace-based disorder, social cohesion, and physical and mental health outcomes of LGBTQIA+ adults in the United States
Place-based disorder and social cohesion may influence lesbian, gay, bisexual, transgender, queer, intersex, asexual, and other sexual and gender minority (LGBTQIA+) adults’ physical and mental health in meaningful ways given their heightened exposure to minority stress and discrimination. However, few studies have examined these associations. Using a sample of 3790 LGBTQIA+ adults in 786 counties from The Population Research in Identity and Disparities for Equality Study, we assessed associations of place-based disorder and social cohesion with depressive symptoms, perceived stress, and physical health using linear mixed models. We tested effect modification by gender modality and ethnoracial group. Residing in counties with higher social cohesion by 1 unit was associated with 1.06 points lower (95% CI, −1.56, −0.56) for depressive symptoms, 1.60 points lower (95% CI, −2.26, −0.94) for perceived stress, and 1.17 points higher (95% CI, 0.43, 1.90) for physical health. Residing in counties with higher place-based disorder by 1 unit was associated with 1.17 points higher (95% CI, 0.32, 2.01) for perceived stress; no association was observed for depressive symptoms or physical health. We did not detect evidence of effect modification by ethnoracial groups, though the association between place-based disorder and depressive symptoms appeared to differ by gender modality. Findings indicate that physical and social environments are important to the health of LGBTQIA+ individuals.
- Open Access
- 26 March 2026
- Published by Oxford University Press (OUP) in American Journal of Epidemiology
Two-way fixed effects models in air pollution epidemiology: a proposed framework for model specifications and recommendations for future studies
Two-way fixed effects (TWFE) models are increasingly applied in air pollution epidemiology. However, it remains unclear how sensitive TWFE models are to different modeling choices and how to transparently select a final model from a wide range of options. This study aims to systematically assess the sensitivity of TWFE models to different fixed effects (FEs) and propose a comprehensive decision-making framework. As an illustrative case study, we investigated associations between weekly ZIP Code Tabulation Area (ZCTA)-level mean PM2.5 concentrations and respiratory hospitalizations in California, 2011–2019. We proposed 52 TWFE models with all plausible spatiotemporal FE combinations and demonstrated a three-stage decision-making framework for final model selection based on permutation tests, equivalence testing, and model complexity. Substantial variations in estimates were observed among models with different combinations of FEs and their interactions. The final model estimated that a 1 μg/m3 increase in weekly mean PM2.5 concentration was associated with a 0.06% increase in weekly respiratory hospitalizations (95% CI, 0.03%, 0.09%; clustered SE by ZCTA). This study highlighted that the choice of FEs in TWFE models can have meaningful influences in the inference of interest. Our proposed decision-making offers a practical guide for future air pollution epidemiological studies using TWFE models.
- 26 March 2026
- Published by Oxford University Press (OUP) in American Journal of Epidemiology
Open AccessTwo-way fixed effects models in air pollution epidemiology: a proposed framework for model specifications and recommendations for future studies
Two-way fixed effects (TWFE) models are increasingly applied in air pollution epidemiology. However, it remains unclear how sensitive TWFE models are to different modeling choices and how to transparently select a final model from a wide range of options. This study aims to systematically assess the sensitivity of TWFE models to different fixed effects (FEs) and propose a comprehensive decision-making framework. As an illustrative case study, we investigated associations between weekly ZIP Code Tabulation Area (ZCTA)-level mean PM2.5 concentrations and respiratory hospitalizations in California, 2011–2019. We proposed 52 TWFE models with all plausible spatiotemporal FE combinations and demonstrated a three-stage decision-making framework for final model selection based on permutation tests, equivalence testing, and model complexity. Substantial variations in estimates were observed among models with different combinations of FEs and their interactions. The final model estimated that a 1 μg/m3 increase in weekly mean PM2.5 concentration was associated with a 0.06% increase in weekly respiratory hospitalizations (95% CI, 0.03%, 0.09%; clustered SE by ZCTA). This study highlighted that the choice of FEs in TWFE models can have meaningful influences in the inference of interest. Our proposed decision-making offers a practical guide for future air pollution epidemiological studies using TWFE models.
- Open Access
Aging Out of the Blue: Region-Specific Epigenetic Clock Calibration for a Blue Zone with the DNAm SuperLearner
Epigenetic clocks estimate biological age from DNA methylation patterns at CpG sites, providing robust predictions of mortality and morbidity risk. “Blue zones”—regions of exceptional longevity—offer a unique opportunity to investigate how biological aging diverges from chronological age. However, standard clocks are typically trained on large, heterogeneous datasets, reflecting average population trends rather than region-specific dynamics. Using data from the Costa Rican Longevity and Healthy Aging Study (CRELES), we profiled DNA methylation from residents of the Nicoya blue zone ( n = 206) and a comparison population in other parts of Costa Rica ( n = 875). We propose training a SuperLearner , an ensemble machine learning approach, on the non-Nicoyan Costa Ricans to optimize predictive performance across existing clocks and flexible machine learners. Theoretically justified by its Oracle property, SuperLearner performs asymptotically as well as the best candidate predictor in the ensemble, resulting in a weighted combination of algorithms used to predict age. We then used this trained model to construct a calibrated hypothesis test comparing residual age distributions between the blue zone region and the comparison population. Comparing our approach to the five top-performing epigenetic clocks (ranked by MSE) in the Costa Rican cohort, only SuperLearner suggested age deceleration (an average of ∼ 1 year) in the non-Nicoyan reference group. Before calibration, SuperLearner showed the strongest evidence for slowed biological aging among blue zone Nicoyans, estimating a three-year reduction ( R̄ 1 = −3.05, 95% CI [−3.64, −2.46]) in epigenetic age. Calibrating with non-Nicoyan Costa Ricans improved consistency between estimates in all clocks, decreasing the estimated aging advantage in Nicoyans to about two years (Δ̂ = −1.96, 95% CI [−2.56, −1.37]). This approach provides a robust framework for estimating longevity in distinct regions when a relevant comparison population is available.
Open AccessAging Out of the Blue: Region-Specific Epigenetic Clock Calibration for a Blue Zone with the DNAm SuperLearner
Epigenetic clocks estimate biological age from DNA methylation patterns at CpG sites, providing robust predictions of mortality and morbidity risk. “Blue zones”—regions of exceptional longevity—offer a unique opportunity to investigate how biological aging diverges from chronological age. However, standard clocks are typically trained on large, heterogeneous datasets, reflecting average population trends rather than region-specific dynamics. Using data from the Costa Rican Longevity and Healthy Aging Study (CRELES), we profiled DNA methylation from residents of the Nicoya blue zone ( n = 206) and a comparison population in other parts of Costa Rica ( n = 875). We propose training a SuperLearner , an ensemble machine learning approach, on the non-Nicoyan Costa Ricans to optimize predictive performance across existing clocks and flexible machine learners. Theoretically justified by its Oracle property, SuperLearner performs asymptotically as well as the best candidate predictor in the ensemble, resulting in a weighted combination of algorithms used to predict age. We then used this trained model to construct a calibrated hypothesis test comparing residual age distributions between the blue zone region and the comparison population. Comparing our approach to the five top-performing epigenetic clocks (ranked by MSE) in the Costa Rican cohort, only SuperLearner suggested age deceleration (an average of ∼ 1 year) in the non-Nicoyan reference group. Before calibration, SuperLearner showed the strongest evidence for slowed biological aging among blue zone Nicoyans, estimating a three-year reduction ( R̄ 1 = −3.05, 95% CI [−3.64, −2.46]) in epigenetic age. Calibrating with non-Nicoyan Costa Ricans improved consistency between estimates in all clocks, decreasing the estimated aging advantage in Nicoyans to about two years (Δ̂ = −1.96, 95% CI [−2.56, −1.37]). This approach provides a robust framework for estimating longevity in distinct regions when a relevant comparison population is available.
- Open Access
- 27 January 2026
- Published by Springer Nature in Research Square
Are blue zones Disappearing?
The historic rise in human longevity among long-lived national populations has slowed down in this century. One proposed explanation is that, as survival to older ages becomes more common, additional gains depend increasingly on reducing mortality at ages where the baseline hazard is dominated by biological aging, making further improvements harder to achieve without interventions that modify aging itself. If this interpretation is correct, then geographically defined longevity hotspots known as blue zones, should show early or pronounced deceleration and/or convergence as other populations catch up. We test this using period vital statistics from three widely cited blue zones: Okinawa (Japan) and in selected municipalities in Sardinia (Italy) and Nicoya (Costa Rica). We compare each blue zone to (i) its national population, and (ii) to three contemporaneous long-lived countries (a composite) with the highest period life expectancy at the time each blue zone was first reported in the scientific literature (Japan, Hong Kong, and France). Using period demographic metrics of longevity (life expectancy at age 70 [e(70)], survival (conditional survival from age 70 to 100), and the prevalence of centenarians from 1991-2019 for Okinawa, for males we found that in the year each was declared a blue zone, e(70) was an average of 9.8% higher relative to composite; conditional survival from ages 70 to 100 was ~2-fold that of composite; and the prevalence of centenarians per 100,000 in Okinawa was 2.7-fold greater than composite. Similar but highly variable patterns were observed for females. All three blue zones were revalidated here as longevity outliers in the year in which they were declared. We also found a broad deceleration in the rate of improvement in these metrics for Okinawa and Nicoya after ~2010, defined as a reduction in the slope of improvement (and in some cases a decline) compared with earlier periods. Overall, the demographic distinctiveness of the three blue zones has narrowed since their initial identification. We attribute this convergence to (1) catch-up among national and other long-lived populations, (2) a general slowdown in late-life mortality improvement consistent with increasing dominance of aging-related causes at advanced ages, and (3) cohort replacement that erodes historically favorable early-life, midlife, and behavioral exposures that characterized early-20th-century cohorts in these regions.
- 27 January 2026
- Published by Springer Nature in Research Square
Open AccessAre blue zones Disappearing?
The historic rise in human longevity among long-lived national populations has slowed down in this century. One proposed explanation is that, as survival to older ages becomes more common, additional gains depend increasingly on reducing mortality at ages where the baseline hazard is dominated by biological aging, making further improvements harder to achieve without interventions that modify aging itself. If this interpretation is correct, then geographically defined longevity hotspots known as blue zones, should show early or pronounced deceleration and/or convergence as other populations catch up. We test this using period vital statistics from three widely cited blue zones: Okinawa (Japan) and in selected municipalities in Sardinia (Italy) and Nicoya (Costa Rica). We compare each blue zone to (i) its national population, and (ii) to three contemporaneous long-lived countries (a composite) with the highest period life expectancy at the time each blue zone was first reported in the scientific literature (Japan, Hong Kong, and France). Using period demographic metrics of longevity (life expectancy at age 70 [e(70)], survival (conditional survival from age 70 to 100), and the prevalence of centenarians from 1991-2019 for Okinawa, for males we found that in the year each was declared a blue zone, e(70) was an average of 9.8% higher relative to composite; conditional survival from ages 70 to 100 was ~2-fold that of composite; and the prevalence of centenarians per 100,000 in Okinawa was 2.7-fold greater than composite. Similar but highly variable patterns were observed for females. All three blue zones were revalidated here as longevity outliers in the year in which they were declared. We also found a broad deceleration in the rate of improvement in these metrics for Okinawa and Nicoya after ~2010, defined as a reduction in the slope of improvement (and in some cases a decline) compared with earlier periods. Overall, the demographic distinctiveness of the three blue zones has narrowed since their initial identification. We attribute this convergence to (1) catch-up among national and other long-lived populations, (2) a general slowdown in late-life mortality improvement consistent with increasing dominance of aging-related causes at advanced ages, and (3) cohort replacement that erodes historically favorable early-life, midlife, and behavioral exposures that characterized early-20th-century cohorts in these regions.
- Open Access
Structured Insight from Unstructured Data: Large Language Models for SDOH-Driven Diabetes Risk Prediction
Social determinants of health (SDOH) play a critical role in Type 2 Diabetes (T2D) management but are often absent from electronic health records and risk prediction models. Most individual-level SDOH data is collected through structured screening tools, which lack the flexibility to capture the complexity of patient experiences and unique needs of a clinic's population. This study explores the use of large language models (LLMs) to extract structured SDOH information from unstructured patient life stories and evaluate the predictive value of both the extracted features and the narratives themselves for assessing diabetes control. We collected unstructured interviews from 65 T2D patients aged 65 and older, focused on their lived experiences, social context, and diabetes management. These narratives were analyzed using LLMs with retrieval-augmented generation to produce concise, actionable qualitative summaries for clinical interpretation and structured quantitative SDOH ratings for risk prediction modeling. The structured SDOH ratings were used independently and in combination with traditional laboratory biomarkers as inputs to linear and tree-based machine learning models (Ridge, Lasso, Random Forest, and XGBoost) to demonstrate how unstructured narrative data can be applied in conventional risk prediction workflows. Finally, we evaluated several LLMs on their ability to predict a patient's level of diabetes control (low, medium, high) directly from interview text with A1C values redacted. LLMs achieved 60% accuracy in predicting diabetes control levels from interview text. This work demonstrates how LLMs can translate unstructured SDOH-related data into structured insights, offering a scalable approach to augment clinical risk models and decision-making.
Open AccessStructured Insight from Unstructured Data: Large Language Models for SDOH-Driven Diabetes Risk Prediction
Social determinants of health (SDOH) play a critical role in Type 2 Diabetes (T2D) management but are often absent from electronic health records and risk prediction models. Most individual-level SDOH data is collected through structured screening tools, which lack the flexibility to capture the complexity of patient experiences and unique needs of a clinic's population. This study explores the use of large language models (LLMs) to extract structured SDOH information from unstructured patient life stories and evaluate the predictive value of both the extracted features and the narratives themselves for assessing diabetes control. We collected unstructured interviews from 65 T2D patients aged 65 and older, focused on their lived experiences, social context, and diabetes management. These narratives were analyzed using LLMs with retrieval-augmented generation to produce concise, actionable qualitative summaries for clinical interpretation and structured quantitative SDOH ratings for risk prediction modeling. The structured SDOH ratings were used independently and in combination with traditional laboratory biomarkers as inputs to linear and tree-based machine learning models (Ridge, Lasso, Random Forest, and XGBoost) to demonstrate how unstructured narrative data can be applied in conventional risk prediction workflows. Finally, we evaluated several LLMs on their ability to predict a patient's level of diabetes control (low, medium, high) directly from interview text with A1C values redacted. LLMs achieved 60% accuracy in predicting diabetes control levels from interview text. This work demonstrates how LLMs can translate unstructured SDOH-related data into structured insights, offering a scalable approach to augment clinical risk models and decision-making.