Ashwin Ramaswamy
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ashwin Ramaswamy has a total of 46 co-authors. Here are the top co-authors by number of shared publications:
- Girish N. Nadkarni
The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, NY, USA
3 shared publications - Alexander William Charney
Icahn School of Medicine at Mount Sinai
2 shared publications - Eric Ng
Department of Urology Weill Cornell Medicine New York, NY United States
2 shared publications - Eyal Klang
BRIDGE GenAI Lab, Beth Israel Deaconess Medical Center, Boston, MA, USA
2 shared publications Hannah Hugo
Department of Medicine, NYC Health + Hospitals / Elmhurst, Icahn School of Medicine at Mount Sinai and Mount Sinai Health System, New York, NY, USA
2 shared publications- Jim Hu
Department of Urology, NewYork-Presbyterian Hospital, Weill Cornell Medicine, New York, NY
2 shared publications
Latest Publications
- Open Access
Ad-verse Effects: Pharmaceutical Advertising Shifts Drug Recommendations by Consumer-Facing AI
Large language models are increasingly used for clinical guidance while their parent companies introduce advertising. We tested whether pharmaceutical ads embedded in the prompts of 12 models from OpenAI, Anthropic, and Google shift drug recommendations across 258,660 API calls and four experiments probing distinct epistemic conditions. When two drugs were both guideline-appropriate, advertising shifted selection of the advertised drug by +12.7 percentage points (P < 0.001), with some model–scenario pairs shifting from 0% to 100%. Google models were the most susceptible (+29.8 pp), followed by OpenAI (+10.9 pp), while Anthropic models showed minimal change (+2.0 pp). When the advertised product lacked evidence or was clinically suboptimal, models resisted. This reveals a structured vulnerability: advertising does not override medical knowledge but fills the space where clinical evidence is underdetermined. An open-response sub-analysis (2,340 calls across three representative models) confirmed that advertising restructures free-text clinical reasoning: models echoed ad claims at 2.7 times the baseline rate while maintaining high stated confidence and rarely disclosing the ad. Susceptibility was provider-dependent (Google: +29.8 pp; OpenAI: +10.9 pp; Anthropic: +2.0 pp). Because this bias operates within clinically correct answers, it is invisible to accuracy-based evaluation, identifying a class of AI safety vulnerability that standard testing cannot detect.
Open AccessAd-verse Effects: Pharmaceutical Advertising Shifts Drug Recommendations by Consumer-Facing AI
Large language models are increasingly used for clinical guidance while their parent companies introduce advertising. We tested whether pharmaceutical ads embedded in the prompts of 12 models from OpenAI, Anthropic, and Google shift drug recommendations across 258,660 API calls and four experiments probing distinct epistemic conditions. When two drugs were both guideline-appropriate, advertising shifted selection of the advertised drug by +12.7 percentage points (P < 0.001), with some model–scenario pairs shifting from 0% to 100%. Google models were the most susceptible (+29.8 pp), followed by OpenAI (+10.9 pp), while Anthropic models showed minimal change (+2.0 pp). When the advertised product lacked evidence or was clinically suboptimal, models resisted. This reveals a structured vulnerability: advertising does not override medical knowledge but fills the space where clinical evidence is underdetermined. An open-response sub-analysis (2,340 calls across three representative models) confirmed that advertising restructures free-text clinical reasoning: models echoed ad claims at 2.7 times the baseline rate while maintaining high stated confidence and rarely disclosing the ad. Susceptibility was provider-dependent (Google: +29.8 pp; OpenAI: +10.9 pp; Anthropic: +2.0 pp). Because this bias operates within clinically correct answers, it is invisible to accuracy-based evaluation, identifying a class of AI safety vulnerability that standard testing cannot detect.
- Open Access
Aggregate benchmark scores obscure patient safety implications of errors across frontier language models
Frontier language models are widely used for health-related queries, yet aggregate benchmark scores do not capture safety implications of errors. We applied the recent Nature Medicine triage benchmark across nine frontier models, comparing directional error profiles, contextual bias, and crisis calibration. In-range accuracy ranged from 75.0% to 87.7%, obscuring clinically meaningful error differences. Looking at directionality of errors, under-triage ranged from 0.0% (GPT-5.2) to 12.3% (GPT-5-mini), over-triage varied independently (9.4–36.9%), and under-triage was uncorrelated with aggregate accuracy. When family members minimized symptoms, all models tested shifted toward lower acuity in ambiguous cases (OR range 2.9–14.9), the only contextual effect observed consistently, and access barriers increased under-triage risk in six. Suicide crisis resource mention rates were low and variable across all models. This cross-model heterogeneity and non-monotonic performance across model generations show that aggregate accuracy alone cannot characterize, rank, or predict the clinical safety of deployed language models.
Open AccessAggregate benchmark scores obscure patient safety implications of errors across frontier language models
Frontier language models are widely used for health-related queries, yet aggregate benchmark scores do not capture safety implications of errors. We applied the recent Nature Medicine triage benchmark across nine frontier models, comparing directional error profiles, contextual bias, and crisis calibration. In-range accuracy ranged from 75.0% to 87.7%, obscuring clinically meaningful error differences. Looking at directionality of errors, under-triage ranged from 0.0% (GPT-5.2) to 12.3% (GPT-5-mini), over-triage varied independently (9.4–36.9%), and under-triage was uncorrelated with aggregate accuracy. When family members minimized symptoms, all models tested shifted toward lower acuity in ambiguous cases (OR range 2.9–14.9), the only contextual effect observed consistently, and access barriers increased under-triage risk in six. Suicide crisis resource mention rates were low and variable across all models. This cross-model heterogeneity and non-monotonic performance across model generations show that aggregate accuracy alone cannot characterize, rank, or predict the clinical safety of deployed language models.
- Top Cited PapersOpen Access
- 23 February 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (5) , 1671-1675
ChatGPT Health performance in a structured test of triage recommendations
ChatGPT Health was launched in January 2026 as OpenAI’s consumer health tool and has reached millions of users. Here we conducted a structured stress test of triage recommendations using 60 clinician-authored vignettes across 21 clinical domains under 16 factorial conditions, yielding 960 total responses. Performance followed an inverted U-shaped pattern, with the most dangerous failures concentrated at clinical extremes—nonurgent presentations (35%) and emergency conditions (48%). Among gold-standard emergencies, the system undertriaged 52% of cases, directing patients with diabetic ketoacidosis or impending respiratory failure to 24–48 h evaluation rather than the emergency department, while correctly triaging classical emergencies such as stroke and anaphylaxis. When family or friends minimized symptoms, indicating anchoring bias, triage recommendations shifted significantly in edge cases (odds ratio = 11.7, 95% confidence interval = 3.7–36.6), with the majority of shifts toward less urgent care. Crisis-intervention messages activated unpredictably across suicidal ideation presentations, occurring more frequently when patients described no specific method than when they did. Patient race, sex and barriers to care did not show significant effects, although confidence intervals did not exclude clinically meaningful differences. These findings reveal missed high-risk emergencies and inconsistent activation of crisis safeguards, raising safety concerns that warrant prospective validation before consumer-scale deployment of artificial intelligence triage systems.
- 23 February 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (5) , 1671-1675
Top Cited PapersOpen AccessChatGPT Health performance in a structured test of triage recommendations
ChatGPT Health was launched in January 2026 as OpenAI’s consumer health tool and has reached millions of users. Here we conducted a structured stress test of triage recommendations using 60 clinician-authored vignettes across 21 clinical domains under 16 factorial conditions, yielding 960 total responses. Performance followed an inverted U-shaped pattern, with the most dangerous failures concentrated at clinical extremes—nonurgent presentations (35%) and emergency conditions (48%). Among gold-standard emergencies, the system undertriaged 52% of cases, directing patients with diabetic ketoacidosis or impending respiratory failure to 24–48 h evaluation rather than the emergency department, while correctly triaging classical emergencies such as stroke and anaphylaxis. When family or friends minimized symptoms, indicating anchoring bias, triage recommendations shifted significantly in edge cases (odds ratio = 11.7, 95% confidence interval = 3.7–36.6), with the majority of shifts toward less urgent care. Crisis-intervention messages activated unpredictably across suicidal ideation presentations, occurring more frequently when patients described no specific method than when they did. Patient race, sex and barriers to care did not show significant effects, although confidence intervals did not exclude clinically meaningful differences. These findings reveal missed high-risk emergencies and inconsistent activation of crisis safeguards, raising safety concerns that warrant prospective validation before consumer-scale deployment of artificial intelligence triage systems.
- Open Access
- 30 December 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (12) , e0314989
Ascertaining provider-level implicit bias in electronic health records with rules-based natural language processing: A pilot study in the case of prostate cancer
Purpose: Implicit, unconscious biases in medicine are personal attitudes about race, ethnicity, gender, and other characteristics that may lead to discriminatory patterns of care. However, there is no consensus on whether implicit bias represents a true predictor of differential care given an absence of real-world studies. We conducted the first real-world pilot study of provider implicit bias by evaluating treatment parity in prostate cancer using unstructured data–the most common way providers document granular details of the patient encounter. Methods and findings: Patients ≥18 years with a diagnosis of very-low to favorable intermediate-risk prostate cancer followed by 3 urologic oncologists from 2010 through 2021. The race Implicit Association Test was administered to all providers. Natural language processing screened human annotation using validated regex ontologies evaluated each provider’s care on four prostate cancer quality indicators: (1) active surveillance utilization; (2) molecular biomarker discussion; (3) urinary function evaluation; and (4) sexual function evaluation. The chi-squared test and phi coefficient were utilized to respectively measure the statistical significance and the strength of association between race and four quality indicators. 1,094 patients were included. While Providers A and B demonstrated no preference on the race Implicit Association Test, Provider C showed preference for White patients. Provider C recommended active surveillance (p<0.01, φ = 0.175) and considered biomarkers (p = 0.047, φ = 0.127) more often in White men than expected, suggestive of treatment imparity. Provider A considered biomarkers (p0.05). Conclusions: In this pilot study, providers’ practice patterns were associated with both patient race and implicit racial preferences in prostate cancer. Alerting providers of existing implicit bias may restore parity, however future assessments are needed to validate this concept.
- 30 December 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (12) , e0314989
Open AccessAscertaining provider-level implicit bias in electronic health records with rules-based natural language processing: A pilot study in the case of prostate cancer
Purpose: Implicit, unconscious biases in medicine are personal attitudes about race, ethnicity, gender, and other characteristics that may lead to discriminatory patterns of care. However, there is no consensus on whether implicit bias represents a true predictor of differential care given an absence of real-world studies. We conducted the first real-world pilot study of provider implicit bias by evaluating treatment parity in prostate cancer using unstructured data–the most common way providers document granular details of the patient encounter. Methods and findings: Patients ≥18 years with a diagnosis of very-low to favorable intermediate-risk prostate cancer followed by 3 urologic oncologists from 2010 through 2021. The race Implicit Association Test was administered to all providers. Natural language processing screened human annotation using validated regex ontologies evaluated each provider’s care on four prostate cancer quality indicators: (1) active surveillance utilization; (2) molecular biomarker discussion; (3) urinary function evaluation; and (4) sexual function evaluation. The chi-squared test and phi coefficient were utilized to respectively measure the statistical significance and the strength of association between race and four quality indicators. 1,094 patients were included. While Providers A and B demonstrated no preference on the race Implicit Association Test, Provider C showed preference for White patients. Provider C recommended active surveillance (p<0.01, φ = 0.175) and considered biomarkers (p = 0.047, φ = 0.127) more often in White men than expected, suggestive of treatment imparity. Provider A considered biomarkers (p0.05). Conclusions: In this pilot study, providers’ practice patterns were associated with both patient race and implicit racial preferences in prostate cancer. Alerting providers of existing implicit bias may restore parity, however future assessments are needed to validate this concept.
- Top Cited PapersOpen Access
- 9 September 2020
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 22 (9) , e20786
Patient Satisfaction With Telemedicine During the COVID-19 Pandemic: Retrospective Cohort Study
Background: New York City was the international epicenter of the COVID-19 pandemic. Health care providers responded by rapidly transitioning from in-person to video consultations. Telemedicine (ie, video visits) is a potentially disruptive innovation; however, little is known about patient satisfaction with this emerging alternative to the traditional clinical encounter. Objective: This study aimed to determine if patient satisfaction differs between video and in-person visits. Methods: In this retrospective observational cohort study, we analyzed 38,609 Press Ganey patient satisfaction survey outcomes from clinic encounters (620 video visits vs 37,989 in-person visits) at a single-institution, urban, quaternary academic medical center in New York City for patients aged 18 years, from April 1, 2019, to March 31, 2020. Time was categorized as pre–COVID-19 and COVID-19 (before vs after March 4, 2020). Wilcoxon-Mann-Whitney tests and multivariable linear regression were used for hypothesis testing and statistical modeling, respectively. Results: We experienced an 8729% increase in video visit utilization during the COVID-19 pandemic compared to the same period last year. Video visit Press Ganey scores were significantly higher than in-person visits (94.9% vs 92.5%; P<.001). In adjusted analyses, video visits (parameter estimate [PE] 2.18; 95% CI 1.20-3.16) and the COVID-19 period (PE 0.55; 95% CI 0.04-1.06) were associated with higher patient satisfaction. Younger age (PE –2.05; 95% CI –2.66 to –1.22), female gender (PE –0.73; 95% CI –0.96 to –0.50), and new visit type (PE –0.75; 95% CI –1.00 to –0.49) were associated with lower patient satisfaction. Conclusions: Patient satisfaction with video visits is high and is not a barrier toward a paradigm shift away from traditional in-person clinic visits. Future research comparing other clinic visit quality indicators is needed to guide and implement the widespread adoption of telemedicine.
- 9 September 2020
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 22 (9) , e20786
Top Cited PapersOpen AccessPatient Satisfaction With Telemedicine During the COVID-19 Pandemic: Retrospective Cohort Study
Background: New York City was the international epicenter of the COVID-19 pandemic. Health care providers responded by rapidly transitioning from in-person to video consultations. Telemedicine (ie, video visits) is a potentially disruptive innovation; however, little is known about patient satisfaction with this emerging alternative to the traditional clinical encounter. Objective: This study aimed to determine if patient satisfaction differs between video and in-person visits. Methods: In this retrospective observational cohort study, we analyzed 38,609 Press Ganey patient satisfaction survey outcomes from clinic encounters (620 video visits vs 37,989 in-person visits) at a single-institution, urban, quaternary academic medical center in New York City for patients aged 18 years, from April 1, 2019, to March 31, 2020. Time was categorized as pre–COVID-19 and COVID-19 (before vs after March 4, 2020). Wilcoxon-Mann-Whitney tests and multivariable linear regression were used for hypothesis testing and statistical modeling, respectively. Results: We experienced an 8729% increase in video visit utilization during the COVID-19 pandemic compared to the same period last year. Video visit Press Ganey scores were significantly higher than in-person visits (94.9% vs 92.5%; P<.001). In adjusted analyses, video visits (parameter estimate [PE] 2.18; 95% CI 1.20-3.16) and the COVID-19 period (PE 0.55; 95% CI 0.04-1.06) were associated with higher patient satisfaction. Younger age (PE –2.05; 95% CI –2.66 to –1.22), female gender (PE –0.73; 95% CI –0.96 to –0.50), and new visit type (PE –0.75; 95% CI –1.00 to –0.49) were associated with lower patient satisfaction. Conclusions: Patient satisfaction with video visits is high and is not a barrier toward a paradigm shift away from traditional in-person clinic visits. Future research comparing other clinic visit quality indicators is needed to guide and implement the widespread adoption of telemedicine.
- Open Access
- 28 May 2020
- Published by JMIR Publications Inc. in JMIR Preprints
Patient Satisfaction With Telemedicine During the COVID-19 Pandemic: Retrospective Cohort Study (Preprint)
BACKGROUND New York City was the international epicenter of the COVID-19 pandemic. Health care providers responded by rapidly transitioning from in-person to video consultations. Telemedicine (ie, video visits) is a potentially disruptive innovation; however, little is known about patient satisfaction with this emerging alternative to the traditional clinical encounter. OBJECTIVE This study aimed to determine if patient satisfaction differs between video and in-person visits. METHODS In this retrospective observational cohort study, we analyzed 38,609 Press Ganey patient satisfaction survey outcomes from clinic encounters (620 video visits vs 37,989 in-person visits) at a single-institution, urban, quaternary academic medical center in New York City for patients aged 18 years, from April 1, 2019, to March 31, 2020. Time was categorized as pre–COVID-19 and COVID-19 (before vs after March 4, 2020). Wilcoxon-Mann-Whitney tests and multivariable linear regression were used for hypothesis testing and statistical modeling, respectively. RESULTS We experienced an 8729% increase in video visit utilization during the COVID-19 pandemic compared to the same period last year. Video visit Press Ganey scores were significantly higher than in-person visits (94.9% vs 92.5%; P<.001). In adjusted analyses, video visits (parameter estimate [PE] 2.18; 95% CI 1.20-3.16) and the COVID-19 period (PE 0.55; 95% CI 0.04-1.06) were associated with higher patient satisfaction. Younger age (PE –2.05; 95% CI –2.66 to –1.22), female gender (PE –0.73; 95% CI –0.96 to –0.50), and new visit type (PE –0.75; 95% CI –1.00 to –0.49) were associated with lower patient satisfaction. CONCLUSIONS Patient satisfaction with video visits is high and is not a barrier toward a paradigm shift away from traditional in-person clinic visits. Future research comparing other clinic visit quality indicators is needed to guide and implement the widespread adoption of telemedicine.
- 28 May 2020
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessPatient Satisfaction With Telemedicine During the COVID-19 Pandemic: Retrospective Cohort Study (Preprint)
BACKGROUND New York City was the international epicenter of the COVID-19 pandemic. Health care providers responded by rapidly transitioning from in-person to video consultations. Telemedicine (ie, video visits) is a potentially disruptive innovation; however, little is known about patient satisfaction with this emerging alternative to the traditional clinical encounter. OBJECTIVE This study aimed to determine if patient satisfaction differs between video and in-person visits. METHODS In this retrospective observational cohort study, we analyzed 38,609 Press Ganey patient satisfaction survey outcomes from clinic encounters (620 video visits vs 37,989 in-person visits) at a single-institution, urban, quaternary academic medical center in New York City for patients aged 18 years, from April 1, 2019, to March 31, 2020. Time was categorized as pre–COVID-19 and COVID-19 (before vs after March 4, 2020). Wilcoxon-Mann-Whitney tests and multivariable linear regression were used for hypothesis testing and statistical modeling, respectively. RESULTS We experienced an 8729% increase in video visit utilization during the COVID-19 pandemic compared to the same period last year. Video visit Press Ganey scores were significantly higher than in-person visits (94.9% vs 92.5%; P<.001). In adjusted analyses, video visits (parameter estimate [PE] 2.18; 95% CI 1.20-3.16) and the COVID-19 period (PE 0.55; 95% CI 0.04-1.06) were associated with higher patient satisfaction. Younger age (PE –2.05; 95% CI –2.66 to –1.22), female gender (PE –0.73; 95% CI –0.96 to –0.50), and new visit type (PE –0.75; 95% CI –1.00 to –0.49) were associated with lower patient satisfaction. CONCLUSIONS Patient satisfaction with video visits is high and is not a barrier toward a paradigm shift away from traditional in-person clinic visits. Future research comparing other clinic visit quality indicators is needed to guide and implement the widespread adoption of telemedicine.