Philip Jarrett
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
Philip Jarrett has a total of 63 co-authors. Here are the top co-authors by number of shared publications:
- Andrew R. Jamieson
Lyda Hill Department of Bioinformatics, UT Southwestern Medical Center, Dallas, USA
6 shared publications Dazhe Cao
Department of Emergency Medicine, The University of Texas Southwestern Medical Center, Dallas, USA
5 shared publicationsAndrew Jamieson
Lyda Hill Department of Bioinformatics, The University of Texas Southwestern Medical Center, Dallas, USA
3 shared publications- Deborah B. Diercks
Department of Emergency Medicine UT Southwestern Medical Center Dallas Texas USA
3 shared publications D. Mark Courtney
Department of Emergency Medicine, University of Texas Southwestern Medical Center, Dallas, USA
2 shared publications- David Hein
Lyda Hill Department of Bioinformatics, UT Southwestern Medical Center, Dallas, USA
2 shared publications
Latest Publications
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
No Mastery without the Machine — Medical Education in the Age of Clinical AI
- 22 October 2026
- Published by Massachusetts Medical Society in NEJM AI
- Vol. 3 (11)
No Mastery without the Machine — Medical Education in the Age of Clinical AI
- Open Access
- 7 October 2026
- Published by Springer Nature in npj Digital Medicine
Prospective deployment of multimodal AI grading for medical student OSCEs
- 7 October 2026
- Published by Springer Nature in npj Digital Medicine
Open AccessProspective deployment of multimodal AI grading for medical student OSCEs
- Open Access
- 19 September 2026
- Published by Springer Nature in npj Digital Medicine
Outcome-grounded effect of clinically stigmatizing information on large language model emergency triage prioritization
- 19 September 2026
- Published by Springer Nature in npj Digital Medicine
Open AccessOutcome-grounded effect of clinically stigmatizing information on large language model emergency triage prioritization
- 21 August 2026
- Published by Elsevier in Academic Radiology
- Vol. 33 (11) , 5122-5131
Triage-Time Machine Learning for Computed Tomography Acquisition Prioritization in the Emergency Department
- 21 August 2026
- Published by Elsevier in Academic Radiology
- Vol. 33 (11) , 5122-5131
Triage-Time Machine Learning for Computed Tomography Acquisition Prioritization in the Emergency Department
- Open Access
- 1 June 2026
- Published by American Medical Association (AMA) in JAMA Network Open
- Vol. 9 (6) , e2617356
Frontier Language Models and Optical Character Recognition Preprocessing Against Invisible Text Injection in AI Peer Review
This study assesses the integration of frontier large language models including tools such as optical character recognition (OCR) into the scientific peer review process to detect invisible text injection in an article under review.
- 1 June 2026
- Published by American Medical Association (AMA) in JAMA Network Open
- Vol. 9 (6) , e2617356
Open AccessFrontier Language Models and Optical Character Recognition Preprocessing Against Invisible Text Injection in AI Peer Review
This study assesses the integration of frontier large language models including tools such as optical character recognition (OCR) into the scientific peer review process to detect invisible text injection in an article under review.
- Open Access
- 14 April 2026
- Published by Elsevier in SSRN Electronic Journal
Frontier Language Models and OCR Preprocessing Mitigate Invisible Text Injection in AI Peer Review
Background: As journals explore artificial intelligence (AI)-assisted peer review, the integrity of AI-generated evaluations has emerged as a critical concern. Invisible text injection (ITI), in which adversarial instructions are hidden from human readers via techniques such as white-on-white text or zero-point font size, has been shown to inflate quality scores assigned by earlier large language models (LLMs). Whether current frontier models remain vulnerable, and whether optical character recognition (OCR) preprocessing can neutralize ITI, has not been established. Methods: Using a published open-access clinical trial as a base manuscript, we created six variants: one unaltered control and five containing a single injected flaw each. Each variant was rendered under four ITI conditions and processed via two text extraction paths: direct PDF extraction (which preserves hidden text) and OCR-based extraction (which does not). Six LLMs served as reviewers: three models utilized in prior ITI research (Claude 3 Haiku, GPT-4o-mini, Gemini 2.0 Flash Lite) and three 2026 frontier models (Claude Sonnet 4.5, GPT-5.2, Gemini 3 Pro), yielding 2,880 total reviews analyzed using a linear mixed-effects model. Results: ITI inflated quality scores for earlier models by 0.82 points on direct extraction (Cohen's d = 1.11; P < .001). Frontier models not only resisted ITI but penalized manuscripts containing it (−0.15 points; d = −0.39; P < .001). OCR preprocessing eliminated the ITI effect entirely across all models and generations. Conclusions: Frontier LLMs and OCR preprocessing represent complementary, implementable defenses against ITI in AI-assisted peer review. Journals adopting AI-augmented review should consider both model selection and preprocessing pipeline as components of editorial integrity.
- 14 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFrontier Language Models and OCR Preprocessing Mitigate Invisible Text Injection in AI Peer Review
Background: As journals explore artificial intelligence (AI)-assisted peer review, the integrity of AI-generated evaluations has emerged as a critical concern. Invisible text injection (ITI), in which adversarial instructions are hidden from human readers via techniques such as white-on-white text or zero-point font size, has been shown to inflate quality scores assigned by earlier large language models (LLMs). Whether current frontier models remain vulnerable, and whether optical character recognition (OCR) preprocessing can neutralize ITI, has not been established. Methods: Using a published open-access clinical trial as a base manuscript, we created six variants: one unaltered control and five containing a single injected flaw each. Each variant was rendered under four ITI conditions and processed via two text extraction paths: direct PDF extraction (which preserves hidden text) and OCR-based extraction (which does not). Six LLMs served as reviewers: three models utilized in prior ITI research (Claude 3 Haiku, GPT-4o-mini, Gemini 2.0 Flash Lite) and three 2026 frontier models (Claude Sonnet 4.5, GPT-5.2, Gemini 3 Pro), yielding 2,880 total reviews analyzed using a linear mixed-effects model. Results: ITI inflated quality scores for earlier models by 0.82 points on direct extraction (Cohen's d = 1.11; P < .001). Frontier models not only resisted ITI but penalized manuscripts containing it (−0.15 points; d = −0.39; P < .001). OCR preprocessing eliminated the ITI effect entirely across all models and generations. Conclusions: Frontier LLMs and OCR preprocessing represent complementary, implementable defenses against ITI in AI-assisted peer review. Journals adopting AI-augmented review should consider both model selection and preprocessing pipeline as components of editorial integrity.
- Open Access
Context-Aware Emergency Department Triage Using Pairwise Comparisons and Bradley-Terry Aggregation
STRUCTURED Objective: To evaluate a ranking approach for emergency department (ED) waiting room prioritization that uses pairwise clinical comparisons aggregated via a Bradley-Terry model, and to assess its cross-site stability without site-specific training. Materials and Methods: Using the Multimodal Clinical Monitoring in the Emergency Department (MC-MED) dataset (118,385 ED visits, Site A), we defined a composite deterioration outcome (intensive care unit [ICU] admission, intubation, vasopressor, ventilation, or death within 6 hours) and evaluated 7 queue-ordering policies across 1,000 simulated shifts. The primary endpoint was Recall@5 for deteriorators; secondary endpoints included area under the receiver operating characteristic curve (AUROC) and simulated time-to-provider (TTP) metrics. External validation used MIMIC-IV-ED (425,087 visits, Site B) with 500 shifts. Methods reported per TRIPOD-LLM. Results: On MC-MED, BT-LLM-Enriched (Bradley-Terry ranking with a large language model [LLM] judge, GPT-4.1, using full diagnoses and medications) exceeded the Emergency Severity Index (ESI) on the primary endpoint: Recall@5 0.587 vs. 0.491 (p<0.001). XGBoost achieved Recall@5 0.648 but required large site-specific labeled training data. On external validation, supervised model performance attenuated (XGBoost AUROC 0.892 to 0.807) while BT-LLM-Enriched remained stable (0.826 to 0.831); the two were statistically indistinguishable on external data. Discussion: Under external validation, supervised model performance attenuated while zero-shot LLM ranking remained stable, suggesting cross-site stability without requiring site-specific training data. Conclusion: Pairwise ranking with an LLM judge significantly outperforms ESI-based ordering and remains stable across sites without local training, matching supervised models on external data.
Open AccessContext-Aware Emergency Department Triage Using Pairwise Comparisons and Bradley-Terry Aggregation
STRUCTURED Objective: To evaluate a ranking approach for emergency department (ED) waiting room prioritization that uses pairwise clinical comparisons aggregated via a Bradley-Terry model, and to assess its cross-site stability without site-specific training. Materials and Methods: Using the Multimodal Clinical Monitoring in the Emergency Department (MC-MED) dataset (118,385 ED visits, Site A), we defined a composite deterioration outcome (intensive care unit [ICU] admission, intubation, vasopressor, ventilation, or death within 6 hours) and evaluated 7 queue-ordering policies across 1,000 simulated shifts. The primary endpoint was Recall@5 for deteriorators; secondary endpoints included area under the receiver operating characteristic curve (AUROC) and simulated time-to-provider (TTP) metrics. External validation used MIMIC-IV-ED (425,087 visits, Site B) with 500 shifts. Methods reported per TRIPOD-LLM. Results: On MC-MED, BT-LLM-Enriched (Bradley-Terry ranking with a large language model [LLM] judge, GPT-4.1, using full diagnoses and medications) exceeded the Emergency Severity Index (ESI) on the primary endpoint: Recall@5 0.587 vs. 0.491 (p<0.001). XGBoost achieved Recall@5 0.648 but required large site-specific labeled training data. On external validation, supervised model performance attenuated (XGBoost AUROC 0.892 to 0.807) while BT-LLM-Enriched remained stable (0.826 to 0.831); the two were statistically indistinguishable on external data. Discussion: Under external validation, supervised model performance attenuated while zero-shot LLM ranking remained stable, suggesting cross-site stability without requiring site-specific training data. Conclusion: Pairwise ranking with an LLM judge significantly outperforms ESI-based ordering and remains stable across sites without local training, matching supervised models on external data.
- Open Access
- 3 March 2026
- Published by Elsevier in Resuscitation Plus
- Vol. 28, 101284
Defining thoracic impedance thresholds for rescue ventilation: a laboratory study
- 3 March 2026
- Published by Elsevier in Resuscitation Plus
- Vol. 28, 101284
Open AccessDefining thoracic impedance thresholds for rescue ventilation: a laboratory study
- Open Access
Automated Assessment of OSCE Physical Exams using Multimodal AI
Background: The assessment of physical examination skills in medical education is resource-intensive and prone to inter-rater variability. While artificial intelligence (AI) has successfully automated the grading of clinical notes and transcripts, evaluating the physical techniques themselves—what students do rather than what they say—remains an unsolved challenge. We evaluated whether a multimodal AI system could assess physical examination skills with expert-level reliability. Methods: In this retrospective ablation study, we analyzed 300 video-recorded encounters from six Objective Structured Clinical Examination (OSCE) stations (cardiovascular, respiratory, gastrointestinal, musculoskeletal, and neurological). We compared the performance of a multimodal AI model (Gemini 2.5 Pro) across single- and multi-camera configurations and isolated input modalities (video, audio-only, transcript-only, visual-only) against standard human grading. The primary outcome was agreement with a physician-adjudicated ground-truth reference standard, measured by quadratic weighted Cohen’s kappa (κ). Results: The AI system using a synchronized 3-camera native video configuration achieved significantly higher reliability (κ = 0.830; 95% CI, 0.773–0.880) than the standard human evaluators (κ = 0.732; 95% CI, 0.687–0.776). Performance followed a strict hierarchy: native video > audio-only > transcript-only > visual-only. Notably, visual-only models failed (κ ≈ 0.20) despite high detection accuracy, revealing a “visual paradox” where models could identify when an action occurred but not how well it was performed without audio cues. Conclusions: A properly configured multimodal AI system can grade physical examination skills with reliability exceeding that of trained human evaluators. Success requires native processing of synchronized audio-visual streams; transcript-based or visual-only approaches are insufficient for high-stakes assessment. These findings suggest that AI can provide scalable, objective, and valid assessment of clinical skills, overcoming the limitations of traditional human grading. Abstract Figure:
Open AccessAutomated Assessment of OSCE Physical Exams using Multimodal AI
Background: The assessment of physical examination skills in medical education is resource-intensive and prone to inter-rater variability. While artificial intelligence (AI) has successfully automated the grading of clinical notes and transcripts, evaluating the physical techniques themselves—what students do rather than what they say—remains an unsolved challenge. We evaluated whether a multimodal AI system could assess physical examination skills with expert-level reliability. Methods: In this retrospective ablation study, we analyzed 300 video-recorded encounters from six Objective Structured Clinical Examination (OSCE) stations (cardiovascular, respiratory, gastrointestinal, musculoskeletal, and neurological). We compared the performance of a multimodal AI model (Gemini 2.5 Pro) across single- and multi-camera configurations and isolated input modalities (video, audio-only, transcript-only, visual-only) against standard human grading. The primary outcome was agreement with a physician-adjudicated ground-truth reference standard, measured by quadratic weighted Cohen’s kappa (κ). Results: The AI system using a synchronized 3-camera native video configuration achieved significantly higher reliability (κ = 0.830; 95% CI, 0.773–0.880) than the standard human evaluators (κ = 0.732; 95% CI, 0.687–0.776). Performance followed a strict hierarchy: native video > audio-only > transcript-only > visual-only. Notably, visual-only models failed (κ ≈ 0.20) despite high detection accuracy, revealing a “visual paradox” where models could identify when an action occurred but not how well it was performed without audio cues. Conclusions: A properly configured multimodal AI system can grade physical examination skills with reliability exceeding that of trained human evaluators. Success requires native processing of synchronized audio-visual streams; transcript-based or visual-only approaches are insufficient for high-stakes assessment. These findings suggest that AI can provide scalable, objective, and valid assessment of clinical skills, overcoming the limitations of traditional human grading. Abstract Figure: