Izzet Turkalp Akbasli
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Izzet Turkalp Akbasli has a total of 74 co-authors. Here are the top co-authors by number of shared publications:
- Oguzhan Serin
Division of Pediatric Emergency Medicine, Department of Pediatrics, Faculty of Medicine, Hacettepe University, Ankara, Türkiye
6 shared publications - Ahmet Fatih Deveci
Department of Health Information Systems, University of Health Sciences, Istanbul, Turkey
3 shared publications - Ahmet Ziya Birbilen
Division of Pediatric Emergency, Department of Pediatrics, Faculty of Medicine, Hacettepe University, Ankara, Turkey
3 shared publications - Animesh Tandon
Department of Heart, Vascular, and Thoracic, Division of Cardiology and Cardiovascular Medicine, Children’s Institute, Cleveland Clinic Children’s, Cleveland, USA
3 shared publications - Muhsin Zahid Ugur
Department of Health Information Systems, University of Health Sciences, Istanbul, Turkey
3 shared publications - Orkun Baloglu
Cleveland Clinic
3 shared publications
Latest Publications
- Open Access
- 11 November 2026
- Published by Taylor & Francis in Future Science OA
- Vol. 12 (1)
The role of digital health technologies in easing the burden of congenital heart disease beyond high-income countries
- 11 November 2026
- Published by Taylor & Francis in Future Science OA
- Vol. 12 (1)
Open AccessThe role of digital health technologies in easing the burden of congenital heart disease beyond high-income countries
- Open Access
- 30 September 2026
- Published by Springer Nature in Research Square
Persistent Cardiac Surgery-Associated Acute Kidney Injury Identifies a High-Risk Phenotype after Pediatric Congenital Heart Surgery
Background Cardiac surgery-associated acute kidney injury (CS-AKI) is heterogeneous, and persistence may define a clinically important postoperative phenotype. We examined early CS-AKI phenotypes, their perioperative associations and clinical outcomes, and exploratory discrimination of persistent CS-AKI. Methods This retrospective single-center cohort included 219 eligible patients from 288 consecutively screened children aged ≤ 18 years who underwent congenital heart surgery in 2023–2024. Early CS-AKI was classified as transient CS-AKI if it had resolved by postoperative day 3 and as persistent CS-AKI if it remained present on postoperative day 3 or if therapeutic renal replacement therapy precluded valid creatinine-based assessment. The primary CatBoost model used preoperative, intraoperative, and first-24-hour postoperative variables while excluding postoperative creatinine and was evaluated using nested cross-validation. Results Early CS-AKI occurred in 69 patients (31.5%): 40 had transient CS-AKI and 29 had persistent CS-AKI. Short-term in-hospital mortality was 0.7%, 5.0%, and 44.8% in the no early CS-AKI, transient CS-AKI, and persistent CS-AKI groups, respectively. In the multivariable model, persistent CS-AKI remained associated with younger age, Society of Thoracic Surgeons-European Association for Cardio-Thoracic Surgery (STAT) category 4–5 procedures, a vasoactive-inotropic score (VIS) > 15, and POD0 fluid accumulation. The primary exploratory model achieved a cross-validated area under the receiver operating characteristic curve of 0.85 (95% confidence interval, 0.77–0.93), with a calibration slope of 0.87 and a Brier score of 0.073. Conclusions Persistent CS-AKI, rather than transient, identified a high-risk postoperative phenotype associated with substantially worse short-term outcomes than transient CS-AKI. Routine perioperative variables showed promising exploratory discrimination of persistence after the first postoperative day, but prospective multicenter external validation is required before clinical use.
- 30 September 2026
- Published by Springer Nature in Research Square
Open AccessPersistent Cardiac Surgery-Associated Acute Kidney Injury Identifies a High-Risk Phenotype after Pediatric Congenital Heart Surgery
Background Cardiac surgery-associated acute kidney injury (CS-AKI) is heterogeneous, and persistence may define a clinically important postoperative phenotype. We examined early CS-AKI phenotypes, their perioperative associations and clinical outcomes, and exploratory discrimination of persistent CS-AKI. Methods This retrospective single-center cohort included 219 eligible patients from 288 consecutively screened children aged ≤ 18 years who underwent congenital heart surgery in 2023–2024. Early CS-AKI was classified as transient CS-AKI if it had resolved by postoperative day 3 and as persistent CS-AKI if it remained present on postoperative day 3 or if therapeutic renal replacement therapy precluded valid creatinine-based assessment. The primary CatBoost model used preoperative, intraoperative, and first-24-hour postoperative variables while excluding postoperative creatinine and was evaluated using nested cross-validation. Results Early CS-AKI occurred in 69 patients (31.5%): 40 had transient CS-AKI and 29 had persistent CS-AKI. Short-term in-hospital mortality was 0.7%, 5.0%, and 44.8% in the no early CS-AKI, transient CS-AKI, and persistent CS-AKI groups, respectively. In the multivariable model, persistent CS-AKI remained associated with younger age, Society of Thoracic Surgeons-European Association for Cardio-Thoracic Surgery (STAT) category 4–5 procedures, a vasoactive-inotropic score (VIS) > 15, and POD0 fluid accumulation. The primary exploratory model achieved a cross-validated area under the receiver operating characteristic curve of 0.85 (95% confidence interval, 0.77–0.93), with a calibration slope of 0.87 and a Brier score of 0.073. Conclusions Persistent CS-AKI, rather than transient, identified a high-risk postoperative phenotype associated with substantially worse short-term outcomes than transient CS-AKI. Routine perioperative variables showed promising exploratory discrimination of persistence after the first postoperative day, but prospective multicenter external validation is required before clinical use.
- 25 September 2026
- Published by Springer Nature in Journal of Medical Systems
- Vol. 50 (1) , 137
Forecast-Driven Dynamic Physician Staffing in a Pediatric Emergency Department: A Prospective Quasi-Experimental Pilot Study
- 25 September 2026
- Published by Springer Nature in Journal of Medical Systems
- Vol. 50 (1) , 137
Forecast-Driven Dynamic Physician Staffing in a Pediatric Emergency Department: A Prospective Quasi-Experimental Pilot Study
- Open Access
- 18 September 2026
- Published by IntechOpen
Machine Learning and High-Fidelity Data in Medicine and Surgery
Artificial intelligence has become increasingly integrated into medicine and surgery, yet much of its clinical impact has historically relied on models trained using low resolution and intermittently sampled data. High fidelity clinical data, defined by continuous acquisition, fine temporal resolution, and direct physiologic measurement, constitute a fundamentally different substrate for machine learning algorithms. Continuous monitoring tools such as near infrared spectroscopy and central venous oxygen saturations provide an uninterrupted stream of information that captures the dynamic behavior of human physiology. These data streams reveal complex patterns of physiologic compensation, adaptation, and eventual failure that remain largely invisible to traditional static clinical datasets. By tracking the delicate balance between oxygen delivery and systemic oxygen extraction in real time, clinicians can better understand the adequacy of perfusion before overt hemodynamic collapse occurs. When paired with machine learning methods capable of modeling temporal dependence, nonlinear interactions, and physiologic coupling, this rich stream of information enables a crucial transition from static risk prediction toward dynamic physiologic understanding. This chapter examines the conceptual foundations, analytic paradigms, data engineering requirements, and specific clinical applications of machine learning using high fidelity data, with particular emphasis on pediatric cardiovascular surgery and cardiac intensive care settings. Key methodologic challenges, interpretability considerations, and the ethical, regulatory, and operational constraints of deploying these computational models at the bedside are thoroughly discussed. The review concludes by highlighting future directions toward real time learning health systems and patient specific digital physiologic models that prioritize the continuous optimization of systemic oxygen delivery and extraction.
- 18 September 2026
- Published by IntechOpen
Open AccessMachine Learning and High-Fidelity Data in Medicine and Surgery
Artificial intelligence has become increasingly integrated into medicine and surgery, yet much of its clinical impact has historically relied on models trained using low resolution and intermittently sampled data. High fidelity clinical data, defined by continuous acquisition, fine temporal resolution, and direct physiologic measurement, constitute a fundamentally different substrate for machine learning algorithms. Continuous monitoring tools such as near infrared spectroscopy and central venous oxygen saturations provide an uninterrupted stream of information that captures the dynamic behavior of human physiology. These data streams reveal complex patterns of physiologic compensation, adaptation, and eventual failure that remain largely invisible to traditional static clinical datasets. By tracking the delicate balance between oxygen delivery and systemic oxygen extraction in real time, clinicians can better understand the adequacy of perfusion before overt hemodynamic collapse occurs. When paired with machine learning methods capable of modeling temporal dependence, nonlinear interactions, and physiologic coupling, this rich stream of information enables a crucial transition from static risk prediction toward dynamic physiologic understanding. This chapter examines the conceptual foundations, analytic paradigms, data engineering requirements, and specific clinical applications of machine learning using high fidelity data, with particular emphasis on pediatric cardiovascular surgery and cardiac intensive care settings. Key methodologic challenges, interpretability considerations, and the ethical, regulatory, and operational constraints of deploying these computational models at the bedside are thoroughly discussed. The review concludes by highlighting future directions toward real time learning health systems and patient specific digital physiologic models that prioritize the continuous optimization of systemic oxygen delivery and extraction.
- Open Access
- 19 August 2026
- Published by IntechOpen
From Problem to Sustainable Change: Implementation-Enhanced Quality Improvement for Pediatric Emergency Care Challenges
Delivering high-quality care in pediatric emergency departments (PEDs) requires more than clinical expertise; it necessitates navigating a high-pressure and high-variability environment compounded by recurring systemic vulnerabilities. This chapter examines key quality challenges in pediatric emergency care, including overcrowding and patient flow, diagnostic accuracy, medication safety, workforce well-being, communication and handoff reliability, and discharge safety. Across these domains, it links epidemiologic burden and clinical relevance to underlying system drivers, measurable indicators, and quality improvement (QI) strategies with demonstrated practical value. However, interventions that demonstrate early effectiveness often show variable sustainability beyond initial implementation. To bridge the gap between theory and practice, the chapter introduces a case-based application of a four-phase framework (Problem Definition, Design, Implementation, and Sustainment and Spread) that integrates traditional QI with implementation science, explicitly clarifying the “who, when, and how” of continuous QI efforts.
- 19 August 2026
- Published by IntechOpen
Open AccessFrom Problem to Sustainable Change: Implementation-Enhanced Quality Improvement for Pediatric Emergency Care Challenges
Delivering high-quality care in pediatric emergency departments (PEDs) requires more than clinical expertise; it necessitates navigating a high-pressure and high-variability environment compounded by recurring systemic vulnerabilities. This chapter examines key quality challenges in pediatric emergency care, including overcrowding and patient flow, diagnostic accuracy, medication safety, workforce well-being, communication and handoff reliability, and discharge safety. Across these domains, it links epidemiologic burden and clinical relevance to underlying system drivers, measurable indicators, and quality improvement (QI) strategies with demonstrated practical value. However, interventions that demonstrate early effectiveness often show variable sustainability beyond initial implementation. To bridge the gap between theory and practice, the chapter introduces a case-based application of a four-phase framework (Problem Definition, Design, Implementation, and Sustainment and Spread) that integrates traditional QI with implementation science, explicitly clarifying the “who, when, and how” of continuous QI efforts.
- 18 July 2026
- Published by Springer Nature in Child's Nervous System
- Vol. 42 (1) , 301
Neurocritical factors associated with mortality and functional recovery in pediatric trauma patients admitted to a tertiary PICU
Purpose To identify early neuroprognostic factors associated with functional neurological outcome in pediatric trauma patients requiring PICU admission. The primary outcome was Glasgow Outcome Scale (GOS) at discharge and 6 months; secondary outcomes were in-hospital mortality and brain death. Methods We conducted a retrospective cohort study of 425 consecutive pediatric trauma admissions to a tertiary PICU from June 2014 to December 2021. Demographics, mechanisms of injury, admission severity scores (GCS, PRISM III, PTS), cranial CT findings, early laboratory markers, and key interventions were analyzed. Factors independently associated with mortality were assessed using multivariable logistic regression, and discrimination was evaluated with receiver operating characteristic analyses. Results Median age was 78 months and 64.2% were male. Falls and traffic accidents were the most common mechanisms. Overall mortality was 7.3% and was predominantly attributable to severe traumatic brain injury (29 of 31 deaths). Brain death occurred in 4.5%. At discharge, 349 of 425 patients (82.1%) had a good neurological outcome (GOS 4–5), whereas 76 (17.9%) had a poor outcome (GOS 1–3), including 31 in-hospital deaths (GOS 1). Among 394 hospital survivors, 6-month follow-up was available for 377 (95.7%); of these, 364 (96.6%) achieved good recovery and 13 (3.4%) had persistent poor outcomes. Among survivors with a poor outcome at discharge and available follow-up, 25 of 38 (65.8%) improved to a good outcome by six months. Seventeen survivors were lost to follow-up. Poor outcome (GOS 1 to 3) was associated with low admission GCS, cerebral edema or midline shift, and early physiologic stress including hyperglycemia and elevated lactate. Factors independently associated with mortality were inotropic support (OR 6.98), blood product administration (OR 6.52), albumin administration (OR 3.69), cerebral edema (OR 3.59), and admission GCS 8 or less (OR 6.82). PRISM III and PTS showed strong discrimination. Conclusion In critically injured children, early neurological compromise and systemic instability that are identifiable during initial stabilization were the strongest factors associated with mortality and functional recovery, supporting neurofocused risk stratification and timely escalation of care.
- 18 July 2026
- Published by Springer Nature in Child's Nervous System
- Vol. 42 (1) , 301
Neurocritical factors associated with mortality and functional recovery in pediatric trauma patients admitted to a tertiary PICU
Purpose To identify early neuroprognostic factors associated with functional neurological outcome in pediatric trauma patients requiring PICU admission. The primary outcome was Glasgow Outcome Scale (GOS) at discharge and 6 months; secondary outcomes were in-hospital mortality and brain death. Methods We conducted a retrospective cohort study of 425 consecutive pediatric trauma admissions to a tertiary PICU from June 2014 to December 2021. Demographics, mechanisms of injury, admission severity scores (GCS, PRISM III, PTS), cranial CT findings, early laboratory markers, and key interventions were analyzed. Factors independently associated with mortality were assessed using multivariable logistic regression, and discrimination was evaluated with receiver operating characteristic analyses. Results Median age was 78 months and 64.2% were male. Falls and traffic accidents were the most common mechanisms. Overall mortality was 7.3% and was predominantly attributable to severe traumatic brain injury (29 of 31 deaths). Brain death occurred in 4.5%. At discharge, 349 of 425 patients (82.1%) had a good neurological outcome (GOS 4–5), whereas 76 (17.9%) had a poor outcome (GOS 1–3), including 31 in-hospital deaths (GOS 1). Among 394 hospital survivors, 6-month follow-up was available for 377 (95.7%); of these, 364 (96.6%) achieved good recovery and 13 (3.4%) had persistent poor outcomes. Among survivors with a poor outcome at discharge and available follow-up, 25 of 38 (65.8%) improved to a good outcome by six months. Seventeen survivors were lost to follow-up. Poor outcome (GOS 1 to 3) was associated with low admission GCS, cerebral edema or midline shift, and early physiologic stress including hyperglycemia and elevated lactate. Factors independently associated with mortality were inotropic support (OR 6.98), blood product administration (OR 6.52), albumin administration (OR 3.69), cerebral edema (OR 3.59), and admission GCS 8 or less (OR 6.82). PRISM III and PTS showed strong discrimination. Conclusion In critically injured children, early neurological compromise and systemic instability that are identifiable during initial stabilization were the strongest factors associated with mortality and functional recovery, supporting neurofocused risk stratification and timely escalation of care.
- 1 July 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (04) , 1-18
OCR-mediated modality dominance in vision–language models: implications for radiology AI trustworthiness
PurposeVision–language models (VLMs) are increasingly proposed for radiologic decision support, yet the security implications of deploying optical character recognition (OCR)-capable models in diagnostic workflows remain poorly characterized. When image-embedded text is not treated as untrusted input, the visual channel becomes vulnerable to adversarial manipulation.ApproachTen VLMs, none validated for clinical diagnosis, were evaluated on 600 brain magnetic resonance imaging studies for binary tumor detection under 5 conditions: clean input, visible radiology report injection, human-imperceptible stealth OCR injection, and multi-stage immune-prompt defense. A total of 27,000 inference calls were analyzed.ResultsAt baseline, performance was heterogeneous, with a median accuracy of 0.69, a sensitivity of 0.79, and a specificity of 0.59. Visible injection caused universal specificity collapse to 0.00 across all models with a false-positive rate (FPR) of 1.00 and a median attack success rate (ASR) of 0.97. Stealth injection, despite being imperceptible to human reviewers, drove substantial degradation with a median accuracy of 0.43, an ASR of 0.57, and an FPR of 0.84. Immune prompting achieved only partial mitigation: under stealth injection, median ASR decreased to 0.44, and accuracy improved to 0.56, yet residual overcalling persisted with a median FPR of 0.67, and three models maintained an FPR of 1.00.ConclusionsCommercial VLMs exhibit a deployment-critical failure mode in radiology-like scenarios: OCR-readable text embedded in images can override pixel-level evidence, even under stealth conditions that evade human inspection. Prompt-level defenses provide insufficient protection. Any clinical integration of VLMs must be governed by system-level safeguards, including OCR-aware input handling, provenance controls, and enforced human verification, before deployment in safety-sensitive environments.
- 1 July 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (04) , 1-18
OCR-mediated modality dominance in vision–language models: implications for radiology AI trustworthiness
PurposeVision–language models (VLMs) are increasingly proposed for radiologic decision support, yet the security implications of deploying optical character recognition (OCR)-capable models in diagnostic workflows remain poorly characterized. When image-embedded text is not treated as untrusted input, the visual channel becomes vulnerable to adversarial manipulation.ApproachTen VLMs, none validated for clinical diagnosis, were evaluated on 600 brain magnetic resonance imaging studies for binary tumor detection under 5 conditions: clean input, visible radiology report injection, human-imperceptible stealth OCR injection, and multi-stage immune-prompt defense. A total of 27,000 inference calls were analyzed.ResultsAt baseline, performance was heterogeneous, with a median accuracy of 0.69, a sensitivity of 0.79, and a specificity of 0.59. Visible injection caused universal specificity collapse to 0.00 across all models with a false-positive rate (FPR) of 1.00 and a median attack success rate (ASR) of 0.97. Stealth injection, despite being imperceptible to human reviewers, drove substantial degradation with a median accuracy of 0.43, an ASR of 0.57, and an FPR of 0.84. Immune prompting achieved only partial mitigation: under stealth injection, median ASR decreased to 0.44, and accuracy improved to 0.56, yet residual overcalling persisted with a median FPR of 0.67, and three models maintained an FPR of 1.00.ConclusionsCommercial VLMs exhibit a deployment-critical failure mode in radiology-like scenarios: OCR-readable text embedded in images can override pixel-level evidence, even under stealth conditions that evade human inspection. Prompt-level defenses provide insufficient protection. Any clinical integration of VLMs must be governed by system-level safeguards, including OCR-aware input handling, provenance controls, and enforced human verification, before deployment in safety-sensitive environments.
- Open Access
- 25 June 2026
- Published by Cambridge University Press (CUP) in Cardiology in the Young
- Vol. 36 (6) , 1178-1186
Preoperative, operative, and immediate postoperative predictors of the onset of low cardiac output syndrome in CHD
Background: Low cardiac output syndrome is a leading cause of morbidity and mortality after CHD surgery. Early, transparent risk estimation at paediatric cardiac ICU admission could guide monitoring and resource allocation. Objectives: To develop and evaluate a multivariable model that estimates the risk of low cardiac output syndrome using routinely available preoperative, intraoperative, and immediate postoperative variables. Methods: In this single-centre retrospective observational cohort, children ≤18 years undergoing CHD surgery between February 2023 and November 2024 were included. Low cardiac output syndrome was defined using prespecified criteria from the Paediatric Cardiac Critical Care Consortium. Candidate predictors were screened in univariate analyses; independent associations were estimated with multivariable logistic regression and backward elimination. Results: Among 191 patients, 46 (24%) developed low cardiac output syndrome. Independent predictors were higher surgical complexity ([RACHS-1] ≥ 4; AOR 3.69, 95% confidence intervals 1.38–9.81), preoperative inotrope use (AOR 2.75, 1.02–7.43), longer cardiopulmonary bypass duration (AOR 1.01 per minute, 1.00–1.02), and a greater number of prior cardiac operations (AOR 2.00 per operation, 1.34–2.97); higher operative weight was protective (AOR 0.92 per kilogram, 0.87–0.97). Model performance metrics were AUROC 0.88 and AUPRC 0.71; at a prespecified decision threshold, accuracy 0.82, positive predictive value 0.66, sensitivity (recall) 0.50, F1 score 0.57, and negative predictive value 0.85. Conclusions: A parsimonious, interpretable model derived from routinely collected data identifies children at increased risk of low cardiac output syndrome at ICU arrival and can inform early intervention and staffing. Prospective multicentre validation and dynamic updating with continuous postoperative physiology are warranted.
- 25 June 2026
- Published by Cambridge University Press (CUP) in Cardiology in the Young
- Vol. 36 (6) , 1178-1186
Open AccessPreoperative, operative, and immediate postoperative predictors of the onset of low cardiac output syndrome in CHD
Background: Low cardiac output syndrome is a leading cause of morbidity and mortality after CHD surgery. Early, transparent risk estimation at paediatric cardiac ICU admission could guide monitoring and resource allocation. Objectives: To develop and evaluate a multivariable model that estimates the risk of low cardiac output syndrome using routinely available preoperative, intraoperative, and immediate postoperative variables. Methods: In this single-centre retrospective observational cohort, children ≤18 years undergoing CHD surgery between February 2023 and November 2024 were included. Low cardiac output syndrome was defined using prespecified criteria from the Paediatric Cardiac Critical Care Consortium. Candidate predictors were screened in univariate analyses; independent associations were estimated with multivariable logistic regression and backward elimination. Results: Among 191 patients, 46 (24%) developed low cardiac output syndrome. Independent predictors were higher surgical complexity ([RACHS-1] ≥ 4; AOR 3.69, 95% confidence intervals 1.38–9.81), preoperative inotrope use (AOR 2.75, 1.02–7.43), longer cardiopulmonary bypass duration (AOR 1.01 per minute, 1.00–1.02), and a greater number of prior cardiac operations (AOR 2.00 per operation, 1.34–2.97); higher operative weight was protective (AOR 0.92 per kilogram, 0.87–0.97). Model performance metrics were AUROC 0.88 and AUPRC 0.71; at a prespecified decision threshold, accuracy 0.82, positive predictive value 0.66, sensitivity (recall) 0.50, F1 score 0.57, and negative predictive value 0.85. Conclusions: A parsimonious, interpretable model derived from routinely collected data identifies children at increased risk of low cardiac output syndrome at ICU arrival and can inform early intervention and staffing. Prospective multicentre validation and dynamic updating with continuous postoperative physiology are warranted.
- 16 April 2026
- Published by Springer Nature in Current Treatment Options in Pediatrics
- Vol. 12 (1) , 16
Artificial Intelligence in Pediatric Cardiac Intensive Care: Clinical Applications, Implementation Challenges, and Future Directions
- 16 April 2026
- Published by Springer Nature in Current Treatment Options in Pediatrics
- Vol. 12 (1) , 16
Artificial Intelligence in Pediatric Cardiac Intensive Care: Clinical Applications, Implementation Challenges, and Future Directions
- Open Access
OCR-Mediated Modality Dominance in Vision-Language Models: Implications for Radiology AI Trustworthiness
1. Background: Vision-language models (VLMs) are increasingly proposed for radiologic decision support, yet the security implications of deploying general-domain, OCR-capable models in diagnostic workflows remain poorly characterized. When image-embedded text is not treated as untrusted input, the visual channel becomes vulnerable to adversarial manipulation through OCR-readable overlays. Methods: Nine commercial VLMs, none intended or validated for clinical diagnosis, were evaluated on 600 brain MRI studies (300 tumor-positive, 300 tumor-negative) for binary tumor detection across four conditions: clean input, visible radiology-report injection, human-imperceptible stealth OCR injection, and a multi-stage immune-prompt defense combining both attack types with enforced visual-priority reasoning. Approximately 27,000 inference calls were analyzed. Primary outcomes included accuracy, attack success rate (ASR), false positive rate (FPR), and masking rate. Results: At baseline, performance was heterogeneous (median accuracy 0.69, sensitivity 0.79, specificity 0.59). Visible injection caused universal specificity collapse (0.00 across all models; FPR 1.00), with a median ASR of 0.97; every model unconditionally privileged the injected text over its own visual analysis. Stealth injection, despite being imperceptible to human reviewers, still drove substantial degradation (median accuracy 0.43; ASR 0.57; FPR 0.84). Immune prompting achieved only partial and inconsistent mitigation: under stealth injection, median ASR decreased to 0.44, and accuracy improved to 0.56, yet residual overcalling persisted (median FPR 0.67), and three models maintained an FPR of 1.00. Conclusions: Commercial VLMs exhibit a deployment-critical failure mode in radiology-like scenarios: OCR-readable text embedded in images can dominate the decision pathway and override pixel-level evidence, even under stealth conditions that evade human inspection. Prompt-level defenses provide insufficient protection. These findings underscore that any clinical integration of VLMs must be gated by system-level safeguards, including OCR-aware input handling, provenance controls, and enforced human verification, before such tools can be considered for safety-sensitive environments.
Open AccessOCR-Mediated Modality Dominance in Vision-Language Models: Implications for Radiology AI Trustworthiness
1. Background: Vision-language models (VLMs) are increasingly proposed for radiologic decision support, yet the security implications of deploying general-domain, OCR-capable models in diagnostic workflows remain poorly characterized. When image-embedded text is not treated as untrusted input, the visual channel becomes vulnerable to adversarial manipulation through OCR-readable overlays. Methods: Nine commercial VLMs, none intended or validated for clinical diagnosis, were evaluated on 600 brain MRI studies (300 tumor-positive, 300 tumor-negative) for binary tumor detection across four conditions: clean input, visible radiology-report injection, human-imperceptible stealth OCR injection, and a multi-stage immune-prompt defense combining both attack types with enforced visual-priority reasoning. Approximately 27,000 inference calls were analyzed. Primary outcomes included accuracy, attack success rate (ASR), false positive rate (FPR), and masking rate. Results: At baseline, performance was heterogeneous (median accuracy 0.69, sensitivity 0.79, specificity 0.59). Visible injection caused universal specificity collapse (0.00 across all models; FPR 1.00), with a median ASR of 0.97; every model unconditionally privileged the injected text over its own visual analysis. Stealth injection, despite being imperceptible to human reviewers, still drove substantial degradation (median accuracy 0.43; ASR 0.57; FPR 0.84). Immune prompting achieved only partial and inconsistent mitigation: under stealth injection, median ASR decreased to 0.44, and accuracy improved to 0.56, yet residual overcalling persisted (median FPR 0.67), and three models maintained an FPR of 1.00. Conclusions: Commercial VLMs exhibit a deployment-critical failure mode in radiology-like scenarios: OCR-readable text embedded in images can dominate the decision pathway and override pixel-level evidence, even under stealth conditions that evade human inspection. Prompt-level defenses provide insufficient protection. These findings underscore that any clinical integration of VLMs must be gated by system-level safeguards, including OCR-aware input handling, provenance controls, and enforced human verification, before such tools can be considered for safety-sensitive environments.