Konrad Samsel
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
Research Subjects
Research Interests
Konrad Samsel has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Zahra Shakeri Hossein Abad
University of Toronto
4 shared publications - Christoffer Dharma
University of Toronto Dalla Lana School of Public Health
2 shared publications - Mohammad Amin Kamaleddin
Temerty Centre for Artificial Intelligence Research and Education in Medicine University of Toronto Toronto M5S 1A8 Ontario Canada
2 shared publications - Abizar Meghani
University of Toronto
1 shared publication - Ahmadreza Ghasemiesfe
Department of Radiology, University of California Davis, Sacramento, CA 95817, USA
1 shared publication - Albert Hsiao
Department of Radiology, University of California San Diego, La Jolla, Calif
1 shared publication
Latest Publications
- Open Access
- 4 September 2026
- Published by Elsevier in Artificial Intelligence in Medicine
- Vol. 182
Too agreeable to be accurate? Sycophancy and diagnostic instability of large language models in medical diagnosis
- 4 September 2026
- Published by Elsevier in Artificial Intelligence in Medicine
- Vol. 182
Open AccessToo agreeable to be accurate? Sycophancy and diagnostic instability of large language models in medical diagnosis
- Open Access
- 11 May 2026
- Published by Elsevier in SSRN Electronic Journal
Too Agreeable to Be Accurate? Sycophancy and Diagnostic Instability of Large Language Models in Medical Diagnosis
Diagnostic accuracy in Large Language Models (LLM) is an increasing concern as physicians employed LLMs into their medical practice. We measured whether diagnostic correctness in LLMs changed after a certainty challenge or clinician specialty framing. To evaluate the performance of LLMs, we curated 120 public clinical vignettes: 40 MultiCaRe-derived clinical narratives and 80 MedMCQA-derived exam style cases. Ten proprietary and open-weight LLMs were tested across four prompt conditions with two passes per condition, yielding 9,600 responses. Diagnostic correctness was evaluated using an LLM-as-a-judge approach. At neutral baseline, accuracy varied by model and was generally lower for MultiCaRe than MedMCQA. GPT-5 had the highest baseline and post-challenge accuracy (75.0% and 76.7%); Claude Sonnet 4 had the highest accuracy changing flip rate (AcFR; 57.5%) and largest post-challenge accuracy loss (-37.5 percentage points). Across all neutral model-case pairs, corrected accuracy decreased from 50.6% to 41.7% after 'Are you sure?' (AcFR = 20.4\%). Correct-to-incorrect transitions (n = 176) exceeded incorrect-to-correct transitions (n = 69). Specialty context prompts produced smaller changes. Adjacent specialty prompts increased overall accuracy by 2.9 percentage points, and differential specialty prompts decreased accuracy by 1.5-1.8 percentage points. The 'Are you sure?' challenge was the strongest source of diagnostic instability in the study; case source also strongly affected accuracy. Diagnostic LLM evaluations should report first pass accuracy, harmful flips, beneficial corrections, and stability under clinically plausible conversational challenges before clinical deployment.
- 11 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessToo Agreeable to Be Accurate? Sycophancy and Diagnostic Instability of Large Language Models in Medical Diagnosis
Diagnostic accuracy in Large Language Models (LLM) is an increasing concern as physicians employed LLMs into their medical practice. We measured whether diagnostic correctness in LLMs changed after a certainty challenge or clinician specialty framing. To evaluate the performance of LLMs, we curated 120 public clinical vignettes: 40 MultiCaRe-derived clinical narratives and 80 MedMCQA-derived exam style cases. Ten proprietary and open-weight LLMs were tested across four prompt conditions with two passes per condition, yielding 9,600 responses. Diagnostic correctness was evaluated using an LLM-as-a-judge approach. At neutral baseline, accuracy varied by model and was generally lower for MultiCaRe than MedMCQA. GPT-5 had the highest baseline and post-challenge accuracy (75.0% and 76.7%); Claude Sonnet 4 had the highest accuracy changing flip rate (AcFR; 57.5%) and largest post-challenge accuracy loss (-37.5 percentage points). Across all neutral model-case pairs, corrected accuracy decreased from 50.6% to 41.7% after 'Are you sure?' (AcFR = 20.4\%). Correct-to-incorrect transitions (n = 176) exceeded incorrect-to-correct transitions (n = 69). Specialty context prompts produced smaller changes. Adjacent specialty prompts increased overall accuracy by 2.9 percentage points, and differential specialty prompts decreased accuracy by 1.5-1.8 percentage points. The 'Are you sure?' challenge was the strongest source of diagnostic instability in the study; case source also strongly affected accuracy. Diagnostic LLM evaluations should report first pass accuracy, harmful flips, beneficial corrections, and stability under clinically plausible conversational challenges before clinical deployment.
- Open Access
- 5 February 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (2) , e0000739
From patient voices to policy: Data analytics reveals patterns in Ontario’s hospital feedback
Patient satisfaction is a central measure of high-performing healthcare systems, yet real-world evaluations at scale remain challenging. In this study, we analyzed 122,194 de-identified patient reviews from 45 Ontario general hospitals between January 2015 and July 2022. We applied a natural language processing (NLP) pipeline using a clinical named entity recognition (NER) model fine-tuned on biomedical literature to extract references to diseases, symptoms, and medical procedures from patient reviews. Geospatial analysis was conducted to examine sentiment patterns based on regional census data related to low-income status and visible-minority composition. Our primary objective was to investigate how the COVID-19 pandemic influenced patient satisfaction trends, with a specific focus on clinical units and hospitals serving marginalized populations. We assessed changes in the proportion of positive comments across time periods and socioeconomic groups using multivariate logistic regression. Our findings show that over 80% of the hospitals studied had fewer than 50% positive reviews, highlighting possible systemic gaps in patient needs. Interestingly, the proportion of negative reviews decreased during the COVID-19 pandemic, suggesting possible changes in patient expectations or increased appreciation for healthcare workers. However, certain units, such as dentistry and radiology, experienced more negative ratings as a proportion of their total reviews. ‘Anxiety’ emerged as a recurrent concern in negative reviews, especially during the start of the pandemic, pointing to the growing awareness of mental health needs. Based on our geospatial analysis, hospitals located in regions with higher percentages of visible minority and low-income populations initially saw higher positive review proportions before COVID-19, but this trend reversed after 2020. Our statistical models confirmed that these shifts were significant, particularly for low-income-serving hospitals. Collectively, these results demonstrate how large-scale unstructured data can identify fundamental drivers of patient satisfaction, while underscoring the urgent need for adaptive strategies to address anxiety and combat systemic inequities.
- 5 February 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (2) , e0000739
Open AccessFrom patient voices to policy: Data analytics reveals patterns in Ontario’s hospital feedback
Patient satisfaction is a central measure of high-performing healthcare systems, yet real-world evaluations at scale remain challenging. In this study, we analyzed 122,194 de-identified patient reviews from 45 Ontario general hospitals between January 2015 and July 2022. We applied a natural language processing (NLP) pipeline using a clinical named entity recognition (NER) model fine-tuned on biomedical literature to extract references to diseases, symptoms, and medical procedures from patient reviews. Geospatial analysis was conducted to examine sentiment patterns based on regional census data related to low-income status and visible-minority composition. Our primary objective was to investigate how the COVID-19 pandemic influenced patient satisfaction trends, with a specific focus on clinical units and hospitals serving marginalized populations. We assessed changes in the proportion of positive comments across time periods and socioeconomic groups using multivariate logistic regression. Our findings show that over 80% of the hospitals studied had fewer than 50% positive reviews, highlighting possible systemic gaps in patient needs. Interestingly, the proportion of negative reviews decreased during the COVID-19 pandemic, suggesting possible changes in patient expectations or increased appreciation for healthcare workers. However, certain units, such as dentistry and radiology, experienced more negative ratings as a proportion of their total reviews. ‘Anxiety’ emerged as a recurrent concern in negative reviews, especially during the start of the pandemic, pointing to the growing awareness of mental health needs. Based on our geospatial analysis, hospitals located in regions with higher percentages of visible minority and low-income populations initially saw higher positive review proportions before COVID-19, but this trend reversed after 2020. Our statistical models confirmed that these shifts were significant, particularly for low-income-serving hospitals. Collectively, these results demonstrate how large-scale unstructured data can identify fundamental drivers of patient satisfaction, while underscoring the urgent need for adaptive strategies to address anxiety and combat systemic inequities.
- 1 November 2025
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 317 (2) , e242961
Generative AI and Foundation Models in Radiology: Applications, Opportunities, and Potential Challenges
Foundation models can streamline imaging workflows, improve diagnostic accuracy, enable personalized treatment, and, via integrated generative artificial intelligence, create synthetic images to augment scarce real-world data; however, successful adoption requires explainability, trust, and workflow integration.
- 1 November 2025
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 317 (2) , e242961
Generative AI and Foundation Models in Radiology: Applications, Opportunities, and Potential Challenges
Foundation models can streamline imaging workflows, improve diagnostic accuracy, enable personalized treatment, and, via integrated generative artificial intelligence, create synthetic images to augment scarce real-world data; however, successful adoption requires explainability, trust, and workflow integration.
- Open Access
- 27 July 2024
- Published by MDPI AG in International Journal of Environmental Research and Public Health
- Vol. 21 (8) , 981
Leukemia Incidence by Occupation and Industry: A Cohort Study of 2.3 Million Workers from Ontario, Canada
Although a significant body of evidence has attributed certain occupational exposures with leukemia, such as benzene, formaldehyde, 1,3-butadiene and ionizing radiation, more research is needed to identify work environments at increased risk for this disease. Our study aimed to identify occupational and industry groups associated with an elevated incidence of leukemia using a diverse cohort of workers’ compensation claimants from Ontario, Canada. A total of 2,363,818 workers in the Occupational Disease Surveillance System (ODSS) cohort, with claims between 1983–2019, were followed for malignant leukemia diagnoses up to 31 December 2019. We used a Cox proportional-hazards model to estimate the relative incidence of leukemia in specific occupation and industry groups. After adjusting for age and birth year, males in protective services (HR = 1.17, 95% CI = 1.02–1.35), metal machining (HR = 1.23, 95% CI = 1.07–1.41), transport (HR = 1.15, 95% CI = 1.06–1.25), and mining occupations (HR = 1.28, 95% CI = 1.02–1.60) had elevated risks of leukemia compared to other workers in the ODSS, with comparable findings by industry. Among female workers, slight risk elevations were observed among product fabricating, assembling, and repairing occupations, with other increased risks seen in furniture and fixture manufacturing, storage, and retail industries. These findings underscore the need for exposure-based studies to better understand occupational hazards in these settings.
- 27 July 2024
- Published by MDPI AG in International Journal of Environmental Research and Public Health
- Vol. 21 (8) , 981
Open AccessLeukemia Incidence by Occupation and Industry: A Cohort Study of 2.3 Million Workers from Ontario, Canada
Although a significant body of evidence has attributed certain occupational exposures with leukemia, such as benzene, formaldehyde, 1,3-butadiene and ionizing radiation, more research is needed to identify work environments at increased risk for this disease. Our study aimed to identify occupational and industry groups associated with an elevated incidence of leukemia using a diverse cohort of workers’ compensation claimants from Ontario, Canada. A total of 2,363,818 workers in the Occupational Disease Surveillance System (ODSS) cohort, with claims between 1983–2019, were followed for malignant leukemia diagnoses up to 31 December 2019. We used a Cox proportional-hazards model to estimate the relative incidence of leukemia in specific occupation and industry groups. After adjusting for age and birth year, males in protective services (HR = 1.17, 95% CI = 1.02–1.35), metal machining (HR = 1.23, 95% CI = 1.07–1.41), transport (HR = 1.15, 95% CI = 1.06–1.25), and mining occupations (HR = 1.28, 95% CI = 1.02–1.60) had elevated risks of leukemia compared to other workers in the ODSS, with comparable findings by industry. Among female workers, slight risk elevations were observed among product fabricating, assembling, and repairing occupations, with other increased risks seen in furniture and fixture manufacturing, storage, and retail industries. These findings underscore the need for exposure-based studies to better understand occupational hazards in these settings.
- Open Access
- 23 July 2021
- Published by University of Toronto Libraries - UOTL in Journal of Undergraduate Life Sciences
- Vol. 15 (1) , 9
The Effects of Commercial Freezing on Vitamin Concentrations in Spinach (Spinacia oleracea)
Commercial food processing has had a significant impact on reducing food spoilage and increasing accessibility to nutrient-dense vegetables. The commercial freezing process, in particular, has given producers the ability to store vegetables with minimized risk of microbial and enzymatic spoilage. Despite the effectiveness of freezing as a preservation method, there is evidence that pre-freezing procedures and prolonged storage can reduce the concentration of vitamins present within certain vegetables. Spinach, one of the most widely produced and consumed vegetables, is particularly susceptible to nutrient loss during the commercial freezing process due to its large surface area and high mineral content. This review summarizes the known effects of the freezing process on hydrophilic and lipophilic vitamins including vitamin C, thiamin, riboflavin, β-carotene, and α-tocopherol. There are two key mechanisms that lead to decreased vitamin concentrations, with the first being attributed to pre-freezing processes including washing and blanching which favours the leaching of hydrophilic vitamins. The second mechanism of vitamin loss is attributed to residual enzymatic activity during storage, where the degree of residual activity can be partially attributed to differences in blanching protocols and freezing practices. Understanding the mechanisms and extent of vitamin loss that the commercial freezing process imparts on leafy green vegetables can help inform future research on improved food processing methods that minimize nutrient loss. Implementing procedures that maintain nutrient retention in frozen vegetables has the potential to assist individuals in achieving their recommended daily intakes of micronutrients.
- 23 July 2021
- Published by University of Toronto Libraries - UOTL in Journal of Undergraduate Life Sciences
- Vol. 15 (1) , 9
Open AccessThe Effects of Commercial Freezing on Vitamin Concentrations in Spinach (Spinacia oleracea)
Commercial food processing has had a significant impact on reducing food spoilage and increasing accessibility to nutrient-dense vegetables. The commercial freezing process, in particular, has given producers the ability to store vegetables with minimized risk of microbial and enzymatic spoilage. Despite the effectiveness of freezing as a preservation method, there is evidence that pre-freezing procedures and prolonged storage can reduce the concentration of vitamins present within certain vegetables. Spinach, one of the most widely produced and consumed vegetables, is particularly susceptible to nutrient loss during the commercial freezing process due to its large surface area and high mineral content. This review summarizes the known effects of the freezing process on hydrophilic and lipophilic vitamins including vitamin C, thiamin, riboflavin, β-carotene, and α-tocopherol. There are two key mechanisms that lead to decreased vitamin concentrations, with the first being attributed to pre-freezing processes including washing and blanching which favours the leaching of hydrophilic vitamins. The second mechanism of vitamin loss is attributed to residual enzymatic activity during storage, where the degree of residual activity can be partially attributed to differences in blanching protocols and freezing practices. Understanding the mechanisms and extent of vitamin loss that the commercial freezing process imparts on leafy green vegetables can help inform future research on improved food processing methods that minimize nutrient loss. Implementing procedures that maintain nutrient retention in frozen vegetables has the potential to assist individuals in achieving their recommended daily intakes of micronutrients.