Felipe Ocampo Osorio
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Felipe Ocampo Osorio has a total of 50 co-authors. Here are the top co-authors by number of shared publications:
Azahar Navarro Beltrán
2 shared publications- Clara Ruiz Torres
Hospital Arnau de Vilanova
2 shared publications - Francisco García-García
Computational Biomedicine Laboratory, Principe Felipe Research Center (CIPF)
2 shared publications - Jesus Alejandro Alzate-Grisales
Unidad Mixta de Imagen Biomédica e Inteligencia Artificial FISABIO-CIPF, Fundación para el Fomento de la Investigación Sanitario y Biomédica de la Comunidad Valenciana, Valencia, Spain
2 shared publications - Jose Daniel Restrepo Orozco
Department of Systems and Informatics, Universidad de Caldas
2 shared publications - Juan Miguel Garcia Gomez
Biomedical Data Science Laboratory (BDSLab), Instituto de Aplicaciones de las Tecnologías de la Información y de las Comunicaciones Avanzadas (ITACA), Universitat Politècnica de València (UPV), Valencia, Spain
2 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by Oxford University Press (OUP) in Oxford Open Digital Health
Implementing AI-CDSS in Low-Resource Maternal Care: Lessons Learned from the Preeclapp Deployment
Hypertensive disorders of pregnancy (HDP) are one of the leading causes of maternal morbidity and mortality in Colombia, particularly in rural regions with limited access to specialized care. In this context, Preeclapp was developed as a clinical decision support system grounded in expert-system logic, designed to provide guideline-based support for maternal care in settings with limited resources and connectivity. This proof-of-concept implementation study analyzes the integration of Preeclapp into clinical practice and the factors that influenced its adoption. The main barriers identified included unstable connectivity, lack of data interoperability, high staff turnover, and initial skepticism toward AI-based technologies. Nevertheless, the project showed preliminary feasibility of use by non-specialized personnel in obstetrics and gynecology, even in low-connectivity environments. To overcome these challenges, the project relied on training sessions, embedded educational resources, continuous technical support, and iterative software adaptations. Our implementation suggests that these strategies were perceived by users as supporting adherence to obstetric protocols, reinforcing clinical decision-making, and enhancing staff confidence, particularly in facilities with limited specialist availability. As a proof-of-concept implementation study, this work did not evaluate diagnostic accuracy or clinical effectiveness; instead, it focused on early implementation feasibility, usability, acceptability, and perceived usefulness. The findings suggest that sustainable adoption of AI-CDSS systems requires strengthening digital literacy and promoting a culture of innovation. Our experience with Preeclapp indicates that adoption and implementation benefited from a modular design, low-connectivity functionality, flexibility, and co-creation with local teams, which helped adapt the tool to each institution’s context.
- 13 November 2026
- Published by Oxford University Press (OUP) in Oxford Open Digital Health
Open AccessImplementing AI-CDSS in Low-Resource Maternal Care: Lessons Learned from the Preeclapp Deployment
Hypertensive disorders of pregnancy (HDP) are one of the leading causes of maternal morbidity and mortality in Colombia, particularly in rural regions with limited access to specialized care. In this context, Preeclapp was developed as a clinical decision support system grounded in expert-system logic, designed to provide guideline-based support for maternal care in settings with limited resources and connectivity. This proof-of-concept implementation study analyzes the integration of Preeclapp into clinical practice and the factors that influenced its adoption. The main barriers identified included unstable connectivity, lack of data interoperability, high staff turnover, and initial skepticism toward AI-based technologies. Nevertheless, the project showed preliminary feasibility of use by non-specialized personnel in obstetrics and gynecology, even in low-connectivity environments. To overcome these challenges, the project relied on training sessions, embedded educational resources, continuous technical support, and iterative software adaptations. Our implementation suggests that these strategies were perceived by users as supporting adherence to obstetric protocols, reinforcing clinical decision-making, and enhancing staff confidence, particularly in facilities with limited specialist availability. As a proof-of-concept implementation study, this work did not evaluate diagnostic accuracy or clinical effectiveness; instead, it focused on early implementation feasibility, usability, acceptability, and perceived usefulness. The findings suggest that sustainable adoption of AI-CDSS systems requires strengthening digital literacy and promoting a culture of innovation. Our experience with Preeclapp indicates that adoption and implementation benefited from a modular design, low-connectivity functionality, flexibility, and co-creation with local teams, which helped adapt the tool to each institution’s context.
- Open Access
- 1 July 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Multimodal Artificial Intelligence for Prostate Cancer Diagnosis and Risk Stratification: A Scoping Review (Preprint)
BACKGROUND Multimodal artificial intelligence (AI) may improve clinically significant prostate cancer (csPCa) detection and risk stratification by integrating imaging, histopathology, molecular, radiomic, and structured clinical data. OBJECTIVE This scoping review mapped recent evidence on multimodal AI for prostate cancer diagnosis, staging, prognosis, and treatment personalisation, with emphasis on data modalities, fusion strategies, validation, reproducibility, and clinical translation. METHODS A PRISMA-ScR-guided search of IEEE Xplore, Scopus, Web of Science, PubMed, and SpringerLink identified English-language peer-reviewed studies published from January 2021 to December 2025. Eligible studies applied AI to at least two data sources for prostate cancer detection, grading, staging, prognosis, recurrence prediction, or treatment selection. Two reviewers screened records and charted clinical task, modalities, fusion strategy, validation design, performance, and reproducibility indicators. RESULTS Twenty-six studies were included. Clinical variables were incorporated in 24 studies, multiparametric magnetic resonance imaging in 18, and whole-slide histopathology in six. Multimodal models usually outperformed unimodal baselines in paired comparisons, including PI-CAI performance above the median radiologist AUROC (0.91 vs 0.86) and trimodal PET/MRI/clinical models reporting AUC values up to 0.955. Evidence was limited by retrospective designs, small cohorts, geographic concentration, scarce public code, inconsistent calibration reporting, and no prospective workflow validation. CONCLUSIONS Multimodal AI is promising for biopsy triage, grading, staging, prognostication, and treatment selection, but current evidence supports research prioritisation rather than routine deployment. Prospective, diverse, transparent, and clinically embedded validation is required before multimodal AI can guide routine prostate cancer decisions.
- 1 July 2026
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessMultimodal Artificial Intelligence for Prostate Cancer Diagnosis and Risk Stratification: A Scoping Review (Preprint)
BACKGROUND Multimodal artificial intelligence (AI) may improve clinically significant prostate cancer (csPCa) detection and risk stratification by integrating imaging, histopathology, molecular, radiomic, and structured clinical data. OBJECTIVE This scoping review mapped recent evidence on multimodal AI for prostate cancer diagnosis, staging, prognosis, and treatment personalisation, with emphasis on data modalities, fusion strategies, validation, reproducibility, and clinical translation. METHODS A PRISMA-ScR-guided search of IEEE Xplore, Scopus, Web of Science, PubMed, and SpringerLink identified English-language peer-reviewed studies published from January 2021 to December 2025. Eligible studies applied AI to at least two data sources for prostate cancer detection, grading, staging, prognosis, recurrence prediction, or treatment selection. Two reviewers screened records and charted clinical task, modalities, fusion strategy, validation design, performance, and reproducibility indicators. RESULTS Twenty-six studies were included. Clinical variables were incorporated in 24 studies, multiparametric magnetic resonance imaging in 18, and whole-slide histopathology in six. Multimodal models usually outperformed unimodal baselines in paired comparisons, including PI-CAI performance above the median radiologist AUROC (0.91 vs 0.86) and trimodal PET/MRI/clinical models reporting AUC values up to 0.955. Evidence was limited by retrospective designs, small cohorts, geographic concentration, scarce public code, inconsistent calibration reporting, and no prospective workflow validation. CONCLUSIONS Multimodal AI is promising for biopsy triage, grading, staging, prognostication, and treatment selection, but current evidence supports research prioritisation rather than routine deployment. Prospective, diverse, transparent, and clinically embedded validation is required before multimodal AI can guide routine prostate cancer decisions.
- Open Access
- 2 June 2026
- Published by Springer Nature in Research Square
From Data Fusion to Decision Support in Prostate Cancer: Evidence Mapping of Multimodal Artificial Intelligence for Diagnosis, Prognosis, and Treatment Selection
Artificial intelligence systems for prostate cancer are increasingly moving beyond single-source image analysis toward models that combine radiology, clinical variables, pathology, molecular measurements, and radiomics. We conducted a PRISMA-ScR-guided evidence map of peer-reviewed English-language studies published from January 2021 to December 2025 and indexed in IEEE Xplore, Scopus, Web of Science, PubMed, or SpringerLink. Eligible studies used AI or machine learning to integrate at least two information sources for prostate cancer detection, grading, staging, prognosis, recurrence prediction, or treatment selection. Two reviewers screened records and extracted study design, clinical task, input modalities, fusion approach, validation strategy, performance, reproducibility, and implementation-relevant reporting. Twenty-six studies met eligibility criteria. Clinical variables were the most frequent non-imaging input (24/26), mpMRI was the dominant imaging source (18/26), and six studies incorporated whole-slide histopathology. Reported paired comparisons usually favoured multimodal approaches, with examples including PI-CAI performance above the median radiologist AUROC (0.91 vs. 0.86) and PET/MRI/clinical models reporting AUC values up to 0.955. However, the evidence remains mostly retrospective, geographically concentrated, weakly reproducible, and rarely calibrated; none of the included studies tested a multimodal AI system prospectively inside a clinical pathway. Current findings support continued development and rigorous evaluation rather than routine deployment. The next phase of research should prioritise prospective pathway studies, diverse external validation, transparent reporting, missing-modality robustness, and patient-centred outcomes.
- 2 June 2026
- Published by Springer Nature in Research Square
Open AccessFrom Data Fusion to Decision Support in Prostate Cancer: Evidence Mapping of Multimodal Artificial Intelligence for Diagnosis, Prognosis, and Treatment Selection
Artificial intelligence systems for prostate cancer are increasingly moving beyond single-source image analysis toward models that combine radiology, clinical variables, pathology, molecular measurements, and radiomics. We conducted a PRISMA-ScR-guided evidence map of peer-reviewed English-language studies published from January 2021 to December 2025 and indexed in IEEE Xplore, Scopus, Web of Science, PubMed, or SpringerLink. Eligible studies used AI or machine learning to integrate at least two information sources for prostate cancer detection, grading, staging, prognosis, recurrence prediction, or treatment selection. Two reviewers screened records and extracted study design, clinical task, input modalities, fusion approach, validation strategy, performance, reproducibility, and implementation-relevant reporting. Twenty-six studies met eligibility criteria. Clinical variables were the most frequent non-imaging input (24/26), mpMRI was the dominant imaging source (18/26), and six studies incorporated whole-slide histopathology. Reported paired comparisons usually favoured multimodal approaches, with examples including PI-CAI performance above the median radiologist AUROC (0.91 vs. 0.86) and PET/MRI/clinical models reporting AUC values up to 0.955. However, the evidence remains mostly retrospective, geographically concentrated, weakly reproducible, and rarely calibrated; none of the included studies tested a multimodal AI system prospectively inside a clinical pathway. Current findings support continued development and rigorous evaluation rather than routine deployment. The next phase of research should prioritise prospective pathway studies, diverse external validation, transparent reporting, missing-modality robustness, and patient-centred outcomes.
- Open Access
ShortKit-ML : an open toolkit for measuring shortcut potential in medical imaging embeddings
Background and Objective: A classifier trained on frozen embeddings can take a shortcut through a sensitive attribute only if the embedding carries information about it beyond what the task label explains: its shortcut potential. We asked whether this can be measured on the embedding alone with a known false positive rate, and whether it predicts how strongly a classifier trained on confounded data relies on the attribute. Methods: ShortKit-ML tests whether an embedding E and a sensitive attribute A are dependent given a task label Y . The primary test is a conditional probe: a logistic regression that predicts A from E and Y , scored by out-of-fold AUROC and calibrated by permuting A within levels of Y , between patients. Five secondary detectors share this null. We audited sex, age and race for seven frozen encoders on CheXpert and four on MIMIC-CXR, and trained heads on confounded CheXpert splits. Results: Under the null the probe rejected in 0.02 of synthetic data sets. With A and Y strongly correlated, an unconditional probe rejected in 0.95 and the conditional probe in 0.03. No detector found signals that leave group means unchanged, unlike probes on squared or product features under the same null. The probe found attribute information in 21 of 21 CheXpert and 12 of 12 MIMIC-CXR cells. Across 42 CheXpert cells the probe score correlated with the true positive rate gap of a confounded head (Spearman 0.60, 95% CI 0.31 to 0.80 resampling encoders), 0.74 within sex and not significantly within age. Three embedding-level mitigations changed the gap by at most 0.01 relative to their controls. Refitting the head on a group-balanced subset closed it. A planted image marker was found in all 10 seeds. Conclusions: One calibrated linear test measures shortcut potential on a frozen embedding, and on CheXpert its score correlated with how strongly a confounded logistic head relied on the attribute, clearly for sex. It does not show that a given head uses the attribute, and lowering the score did not remove reliance, so the head itself must be tested as well. Code is at https://github.com/criticaldata/ShortKit-ML .
Open AccessShortKit-ML : an open toolkit for measuring shortcut potential in medical imaging embeddings
Background and Objective: A classifier trained on frozen embeddings can take a shortcut through a sensitive attribute only if the embedding carries information about it beyond what the task label explains: its shortcut potential. We asked whether this can be measured on the embedding alone with a known false positive rate, and whether it predicts how strongly a classifier trained on confounded data relies on the attribute. Methods: ShortKit-ML tests whether an embedding E and a sensitive attribute A are dependent given a task label Y . The primary test is a conditional probe: a logistic regression that predicts A from E and Y , scored by out-of-fold AUROC and calibrated by permuting A within levels of Y , between patients. Five secondary detectors share this null. We audited sex, age and race for seven frozen encoders on CheXpert and four on MIMIC-CXR, and trained heads on confounded CheXpert splits. Results: Under the null the probe rejected in 0.02 of synthetic data sets. With A and Y strongly correlated, an unconditional probe rejected in 0.95 and the conditional probe in 0.03. No detector found signals that leave group means unchanged, unlike probes on squared or product features under the same null. The probe found attribute information in 21 of 21 CheXpert and 12 of 12 MIMIC-CXR cells. Across 42 CheXpert cells the probe score correlated with the true positive rate gap of a confounded head (Spearman 0.60, 95% CI 0.31 to 0.80 resampling encoders), 0.74 within sex and not significantly within age. Three embedding-level mitigations changed the gap by at most 0.01 relative to their controls. Refitting the head on a group-balanced subset closed it. A planted image marker was found in all 10 seeds. Conclusions: One calibrated linear test measures shortcut potential on a frozen embedding, and on CheXpert its score correlated with how strongly a confounded logistic head relied on the attribute, clearly for sex. It does not show that a given head uses the attribute, and lowering the score did not remove reliance, so the head itself must be tested as well. Code is at https://github.com/criticaldata/ShortKit-ML .
- Open Access
Developing and Testing an Engineering Framework for Curiosity-Driven and Humble AI in Clinical Decision Support
Background: We present BODHI (Balanced, Open-minded, Diagnostic, Humble, and Inquisitive), an engineering framework for curiosity-driven and humble clinical decision support AI. Despite growing capabilities, large language models (LLMs) often express inappropriate confidence, conflating statistical pattern recognition with genuine medical understanding. BODHI addresses this through a dual-reflective architecture that: (1) decomposes epistemic uncertainty into task-specific dimensions, and (2) constrains model responses using virtue-based stance rules derived from a Virtue Activation Matrix. Methods: We validate the framework through controlled evaluation on 200 clinical vignettes from HealthBench Hard, assessing GPT-4o-mini and GPT-4.1-mini across 5 random seeds (1,800 total observations). Statistical analysis included bootstrap resampling, paired t-tests, and effect size computation (Supplementary Materials S3) Findings: BODHI significantly improved overall clinical response quality (GPT-4.1-mini: +17.3pp, p < 0.0001, Cohen’s d = 0.50; GPT-4o-mini: +7.4pp, p < 0.0001, Cohen’s d = 0.22) while achieving very large effect sizes on curiosity (context-seeking rate: Cohen’s d = 16.38 and 19.54) and humility (hedging: d = 5.80 for GPT-4.1-mini) metrics. Crucially, 97.3% of GPT-4.1-mini responses and 73.5% of GPT-4o-mini responses included appropriate clarifying questions, compared to 7.8% and 0.0% at baseline, demonstrating the framework’s effectiveness in eliciting information-gathering behavior. Interpretation: These findings suggest LLMs can be reliably constrained to operate within epistemic boundaries when provided with structured uncertainty decomposition and virtue-aligned response rules, offering a pathway toward safer clinical AI deployment.
Open AccessDeveloping and Testing an Engineering Framework for Curiosity-Driven and Humble AI in Clinical Decision Support
Background: We present BODHI (Balanced, Open-minded, Diagnostic, Humble, and Inquisitive), an engineering framework for curiosity-driven and humble clinical decision support AI. Despite growing capabilities, large language models (LLMs) often express inappropriate confidence, conflating statistical pattern recognition with genuine medical understanding. BODHI addresses this through a dual-reflective architecture that: (1) decomposes epistemic uncertainty into task-specific dimensions, and (2) constrains model responses using virtue-based stance rules derived from a Virtue Activation Matrix. Methods: We validate the framework through controlled evaluation on 200 clinical vignettes from HealthBench Hard, assessing GPT-4o-mini and GPT-4.1-mini across 5 random seeds (1,800 total observations). Statistical analysis included bootstrap resampling, paired t-tests, and effect size computation (Supplementary Materials S3) Findings: BODHI significantly improved overall clinical response quality (GPT-4.1-mini: +17.3pp, p < 0.0001, Cohen’s d = 0.50; GPT-4o-mini: +7.4pp, p < 0.0001, Cohen’s d = 0.22) while achieving very large effect sizes on curiosity (context-seeking rate: Cohen’s d = 16.38 and 19.54) and humility (hedging: d = 5.80 for GPT-4.1-mini) metrics. Crucially, 97.3% of GPT-4.1-mini responses and 73.5% of GPT-4o-mini responses included appropriate clarifying questions, compared to 7.8% and 0.0% at baseline, demonstrating the framework’s effectiveness in eliciting information-gathering behavior. Interpretation: These findings suggest LLMs can be reliably constrained to operate within epistemic boundaries when provided with structured uncertainty decomposition and virtue-aligned response rules, offering a pathway toward safer clinical AI deployment.