Raphael T. Malikian, MBBS, BSc (Hons) translates healthcare AI research into practical, clinically grounded questions for builders, clinicians, researchers, and governance teams.

GitHub: https://github.com/rtmalikian

LinkedIn: http://www.linkedin.com/in/raphael-t-malikian-mbbs-bsc-hons-71075436a

---

Healthcare AI Daily translates one recent peer-reviewed healthcare AI article into a short practical briefing. Today: a locally-run large language model matches fine-tuned approaches for detecting self-harm in NHS electronic mental health records — without sending patient data anywhere. Privacy-preserving AI for the most sensitive clinical data. We explain the F1 score, what the NHS is, and the balanced limitations of this single-site study.

Source article

Title: Detection of Self-Harm in Electronic Mental Health Records Using Privacy-Preserving Local Language Models: Methodological Study

Authors: Andrey Kormilitzin, Dan W Joyce, Apostolos Tsiachristas, Rohan Borschmann, Navneet Kapur, Galit Geulayov

Journal: JMIR Mental Health

Published: 2 June 2026

DOI: https://doi.org/10.2196/87586

Article: https://mental.jmir.org/2026/1/e87586

PDF: https://mental.jmir.org/2026/1/e87586/pdf

PubMed PMID: 42227874

Key reported results: 1,000 patients, 1,352 annotated notes from Oxford Health NHS Foundation Trust; Gemma3-27b (local, no data leaves institution) vs fine-tuned RoBERTa; self-harm detection F1=0.92 (both models); global weighted F1: Gemma3-27b=0.88 vs RoBERTa=0.85; recent self-harm (51 notes): Gemma3-27b F1=0.79; prompt engineering only, no gradient-based fine-tuning; privacy-preserving: model runs entirely within NHS secure infrastructure.

This video is educational commentary, not medical advice. Source screenshots and figures are used for attributed research discussion.

⚠️ Content disclaimer: This episode discusses self-harm detection in clinical records. If you or someone you know is struggling with thoughts of self-harm, please reach out for help immediately. In the United States, call or text 988 (Suicide and Crisis Lifeline). In the United Kingdom, call 116 123 (Samaritans). Things can get better, and you do not have to face this alone.

Synthetic voice disclosure: narration generated with edge-tts en-US-AndrewNeural and post-processed for clarity.

Watch this episode on YouTube: https://youtu.be/E9EPRWN89LM

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Raphael T. Malikian, MBBS, BSc (Hons) translates healthcare AI research into practical, clinically grounded questions for builders, clinicians, researchers, and governance teams.

GitHub: https://github.com/rtmalikian

LinkedIn: http://www.linkedin.com/in/raphael-t-malikian-mbbs-bsc-hons-71075436a

---

Healthcare AI Daily translates one recent peer-reviewed healthcare AI article into a short practical briefing. Today: a cross-sectional survey reveals that radiologists using an FDA-cleared AI tool for intracranial hemorrhage detection report conditional trust, substantial false-positive burden, and limited workflow savings — with neuroradiologists especially critical.

Source article

Title: Radiologist Perceptions of an AI Tool for Intracranial Hemorrhage Detection in Teleradiology: Cross-Sectional Survey Study

Authors: Andrew J Del Gaizo, Jackson K Del Gaizo, Troy A Shahoumian

Journal: JMIR Human Factors

Published: 2 June 2026

DOI: https://doi.org/10.2196/92145

Article: https://humanfactors.jmir.org/2026/1/e92145

PDF: https://www.jmir.org/2026/1/e92145/PDF

PubMed PMID: 42228936

Key reported results: 65 radiologists surveyed; only 18.5% found false-positive alerts acceptable; 50.8% trust AI when it agrees vs 3.1% when it conflicts; 32.3% agreed AI correctly identifies most ICH cases; 33.8% said FP review time outweighs benefits (primary endpoint); only 10.8% reported reduced interpretation time; neuroradiologists more critical (52.2% vs 23.8%); 32.3% worried about legal vulnerability when overriding AI.

This video is educational commentary, not medical advice. Source screenshots and figures are used for attributed research discussion. Synthetic voice disclosure: narration generated with edge-tts en-US-AndrewNeural and post-processed for clarity.

Watch this episode on YouTube: https://youtu.be/fgWL0JxrCnU

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Raphael T. Malikian, MBBS, BSc (Hons) translates healthcare AI research into practical, clinically grounded questions for builders, clinicians, researchers, and governance teams.

GitHub: https://github.com/rtmalikian

LinkedIn: http://www.linkedin.com/in/raphael-t-malikian-mbbs-bsc-hons-71075436a

---

Healthcare AI Daily translates one recent peer-reviewed healthcare AI article into a short practical briefing. Today: why newer reasoning large language models can still generate stereotyped medical vignettes — and why healthcare teams should monitor representational bias before using synthetic cases in education, evaluation, or workflow design.

Source article

Title: Evaluating the Potential of Reasoning Large Language Models to Perpetuate Racial and Gender Disease Stereotypes in Health Care

Authors: Joshua J Docking, Lee X Li, Bradley D Menz, Stephen Bacchi, Ashley M Hopkins, Michael J Sorich

Journal: Journal of Medical Internet Research

Published: 28 May 2026

DOI: https://doi.org/10.2196/82256

Article: https://www.jmir.org/2026/1/e82256

PDF: https://www.jmir.org/2026/1/e82256/PDF

PubMed PMID: 42208042; PMCID: PMC13218561

Key reported results: 36,000 generated clinical vignettes; o3-mini had over 20% racial misrepresentation in 14/18 conditions; DeepSeek-R1 in 16/18; gender misrepresentation crossed the threshold in 10/18 and 12/18 conditions respectively; 16/20 sampled DeepSeek-R1 reasoning traces invoked disease-demographic associations.

This video is educational commentary, not medical advice. Source screenshots and figures are used for attributed research discussion. Synthetic voice disclosure: narration generated with edge-tts en-US-AndrewNeural and post-processed for clarity.

Watch this episode on YouTube: https://youtu.be/O9vRJFfWJGQ

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Raphael T. Malikian, MBBS, BSc (Hons) translates healthcare AI research into practical, clinically grounded questions for builders, clinicians, researchers, and governance teams.

GitHub: https://github.com/rtmalikian

LinkedIn: http://www.linkedin.com/in/raphael-t-malikian-mbbs-bsc-hons-71075436a

---

Healthcare AI Daily translates one healthcare AI paper into a short practical briefing for builders, clinicians, researchers, and governance teams. Today: why medical AI models can look accurate while learning the wrong shortcut — and how dataset bias audits can help teams find risks before deployment.

Source article

Title: Detecting dataset bias in medical AI using a generalized and modality agnostic auditing approach

Authors: Nathan Drenkow, Mitchell Pavlak, Keith Harrigian, Ayah Zirikly, Adarsh Subbaswamy, Mohammad Mehdi Farhangi, Nicholas Petrick, Mathias Unberath

Journal: npj Digital Medicine

Published: 29 May 2026

DOI: https://doi.org/10.1038/s41746-026-02807-y

Article: https://www.nature.com/articles/s41746-026-02807-y

Publisher supplementary PDF: https://static-content.springer.com/esm/art%3A10.1038%2Fs41746-026-02807-y/MediaObjects/41746_2026_2807_MOESM1_ESM.pdf

Keywords: healthcare AI, medical AI, dataset bias, shortcut learning, AI governance, model validation, clinical artificial intelligence, machine learning bias, G-AUDIT, npj Digital Medicine, FDA, external validation, subgroup analysis, drift monitoring.

This video is educational commentary, not medical advice. Source screenshots and supplementary figures are used for attributed research discussion.

Like, subscribe, and enable notifications for daily Healthcare AI episodes. Healthcare AI Weekly releases every Friday at 9 AM.

Watch this episode on YouTube: https://youtu.be/ejxWoBr2upk

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Mental-R1: Can we teach AI to reason like a clinician for mental health assessment?

Researchers at Oxford developed CRPO (Cognitive Relative Policy Optimization), a reinforcement learning framework that aligns LLM reasoning with human cognitive processes for mental health assessment. Mental-R1 outperforms GPT-5, DeepSeek-R1, and domain-specific models across 8 mental health datasets covering depression, anxiety, suicide risk, stress, and loneliness.

Key findings:

- 80.8% accuracy on depression severity classification

- 10.4 percentage point average improvement in weighted F1-score

- Human-like reasoning pattern: broad exploration followed by confident decisions

- Outperforms GPT-5, DeepSeek-R1, GPT-4o, and all domain-specific models

Paper: arxiv.org/abs/2606.13176

Authors: Xin Wang, Boyan Gao, Yibo Yang, David A. Clifton

#HealthcareAI #MentalHealth #ClinicalAI #MachineLearning #ReinforcementLearning #DigitalHealth #AIResearch #LargeLanguageModels #MentalHealthAI #AIinMedicine #HealthTech

MEDICAL DISCLAIMER: This podcast is for educational and informational purposes only. It is not intended to diagnose, treat, cure, or prevent any medical condition, and should not be relied upon as medical advice. The content discussed reflects published research and does not constitute clinical recommendations. If you have any health concerns or medical questions, please consult a licensed healthcare professional immediately.

In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Created by Raphael T. Malikian, MBBS, BSc (Hons)

For comments, questions, or concerns: [email protected]

YouTube: https://www.youtube.com/@RaphaelMalikian-g4h

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

New AI approaches to HER2 scoring in breast cancer pathology slides, improving consistency and speed of this critical biomarker assessment.

Watch this episode on YouTube: https://youtu.be/ydhDqkzKOgY

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Examining how well LLMs calibrate their confidence when interpreting ultrasound images, and what this means for clinical decision support systems.

Watch this episode on YouTube: https://youtu.be/NuRwJ8dPiJ4

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.

Exploring how LLMs are being used in professional development across healthcare, with implications for training, continuing education, and clinical competency assessment.

Watch this episode on YouTube: https://youtu.be/QVpKbIyXdbk

YouTube Channel: https://www.youtube.com/@RaphaelMalikian-g4h

Created by Raphael T. Malikian ([email protected]). In true AI fashion, this podcast was created with AI tools including text-to-speech using Microsoft Edge TTS and Hermes Agent by Nous Research.