# HAI-DEF

**URL:** https://discuss.ai.google.dev/c/hai-def/62.md

[Latest](https://discuss.ai.google.dev/latest.md) · [Categories](https://discuss.ai.google.dev/categories.md) · [Tags](https://discuss.ai.google.dev/tags.md)

---

## [About the Health AI Developer Foundations (HAI-DEF) forum](https://discuss.ai.google.dev/t/about-the-health-ai-developer-foundations-hai-def-forum/85613)

<div class="topic-metadata">

**Author:** [@kristen](https://discuss.ai.google.dev/u/kristen)\
**Replies:** 13\
**Last updated:** [October 10, 2026, 9:27pm UTC](https://discuss.ai.google.dev/t/about-the-health-ai-developer-foundations-hai-def-forum/85613 "2026-10-10T21:27:34Z")

</div>

Discussion around Health AI Developer Foundations (HAI-DEF) open models. Get started adapting models for health applications: Find models on Hugging Face Get started with documentation Join the community and start co…

---

## [MedGemma 1.5 4B on vLLM with guided JSON: output pads with whitespace until max\_new\_tokens instead of finishing](https://discuss.ai.google.dev/t/medgemma-1-5-4b-on-vllm-with-guided-json-output-pads-with-whitespace-until-max-new-tokens-instead-of-finishing/186784)

<div class="topic-metadata">

**Author:** [@Aravind\_S](https://discuss.ai.google.dev/u/Aravind_S)\
**Replies:** 7\
**Last updated:** [November 17, 2026, 1:50pm UTC](https://discuss.ai.google.dev/t/medgemma-1-5-4b-on-vllm-with-guided-json-output-pads-with-whitespace-until-max-new-tokens-instead-of-finishing/186784 "2026-11-17T13:50:55Z")

</div>

Hi everyone, We’re building a healthcare application that uses MedGemma to extract structured information from medical documents into JSON. We self-host the model and have run into two problems with long structured outp…

---

## [Challenges Building Reliable AI Applications for Healthcare Use Cases](https://discuss.ai.google.dev/t/challenges-building-reliable-ai-applications-for-healthcare-use-cases/183694)

<div class="topic-metadata">

**Author:** [@codeautomation](https://discuss.ai.google.dev/u/codeautomation)\
**Replies:** 3\
**Last updated:** [November 11, 2026, 7:28am UTC](https://discuss.ai.google.dev/t/challenges-building-reliable-ai-applications-for-healthcare-use-cases/183694 "2026-11-11T07:28:02Z")

</div>

Hi HAI-DEF community, I have been exploring how AI models can support healthcare applications and learning about the challenges involved in building reliable health AI systems. Healthcare AI introduces unique requireme…

---

## [Seeking GPU Compute Resources or Cloud Credits for Fine-Tuning MedGemma 4B for Uzbek and Russian Healthcare](https://discuss.ai.google.dev/t/seeking-gpu-compute-resources-or-cloud-credits-for-fine-tuning-medgemma-4b-for-uzbek-and-russian-healthcare/188681)

<div class="topic-metadata">

**Author:** [@Mashxuraxon\_Mamatxal](https://discuss.ai.google.dev/u/Mashxuraxon_Mamatxal)\
**Replies:** 1\
**Last updated:** [October 16, 2026, 8:14am UTC](https://discuss.ai.google.dev/t/seeking-gpu-compute-resources-or-cloud-credits-for-fine-tuning-medgemma-4b-for-uzbek-and-russian-healthcare/188681 "2026-10-16T08:14:40Z")

</div>

Hello HAI-DEF Community, My name is Mashxuraxon, and I am the founder of Clinora.AI, an early-stage healthcare AI startup based in Uzbekistan. We are working on adapting Google MedGemma 4B-IT to better support Uzbek (L…

---

## [Best practices for preparing healthcare AI pipelines](https://discuss.ai.google.dev/t/best-practices-for-preparing-healthcare-ai-pipelines/184195)

<div class="topic-metadata">

**Author:** [@Sonali\_Kumari1](https://discuss.ai.google.dev/u/Sonali_Kumari1)\
**Replies:** 1\
**Last updated:** [October 14, 2026, 4:07am UTC](https://discuss.ai.google.dev/t/best-practices-for-preparing-healthcare-ai-pipelines/184195 "2026-10-14T04:07:18Z")

</div>

Hi everyone, Did you catch our latest HAI-DEF newsletter that went out on Sept 21? In this edition, we highlighted real-world community deployments, shared concrete ways for you to help steer our roadmap, and previewed…

---

## [ATRIA | EchoTrace: Fine-Tuning MedGemma 1.5 for Polygon-Based Heart Structure Contouring](https://discuss.ai.google.dev/t/atria-echotrace-fine-tuning-medgemma-1-5-for-polygon-based-heart-structure-contouring/172907)

<div class="topic-metadata">

**Author:** [@ShehabAnwer](https://discuss.ai.google.dev/u/ShehabAnwer)\
**Replies:** 4\
**Last updated:** [October 11, 2026, 2:30pm UTC](https://discuss.ai.google.dev/t/atria-echotrace-fine-tuning-medgemma-1-5-for-polygon-based-heart-structure-contouring/172907 "2026-10-11T14:30:42Z")

</div>

Esteemed HAI-DEF & MedGemma Team, I’m Shehab Anwer, a cardiologist. I recently shared some of my progress with Rory Pilgrim, who kindly suggested I post here to connect with the right folks for your valuable insights an…

---

## [Hausa Healthcare LLM Preference & Safety Evaluation Dataset](https://discuss.ai.google.dev/t/hausa-healthcare-llm-preference-safety-evaluation-dataset/183601)

<div class="topic-metadata">

**Author:** [@ABDULFATAH\_ADAM\_ABDU](https://discuss.ai.google.dev/u/ABDULFATAH_ADAM_ABDU)\
**Replies:** 3\
**Last updated:** [September 21, 2026, 2:12pm UTC](https://discuss.ai.google.dev/t/hausa-healthcare-llm-preference-safety-evaluation-dataset/183601 "2026-09-21T14:12:11Z")

</div>

Hello HAI-DEF community, I am a computational linguist and Hausa NLP/AI data specialist working on a Hausa LLM Healthcare & Safety Preference Evaluation Dataset. The dataset is developed collaboratively by computationa…

---

## [MedGemma-4B 1.5 for structured lab-report extraction: quantization reliability, latency on-prem, and validation guidance](https://discuss.ai.google.dev/t/medgemma-4b-1-5-for-structured-lab-report-extraction-quantization-reliability-latency-on-prem-and-validation-guidance/177891)

<div class="topic-metadata">

**Author:** [@kkdu](https://discuss.ai.google.dev/u/kkdu)\
**Replies:** 4\
**Last updated:** [August 21, 2026, 12:28pm UTC](https://discuss.ai.google.dev/t/medgemma-4b-1-5-for-structured-lab-report-extraction-quantization-reliability-latency-on-prem-and-validation-guidance/177891 "2026-08-21T12:28:32Z")

</div>

Hi Medgemma team, We’re building a local, on-prem clinical decision-support pipeline that extracts structured findings from lab/pathology reports and grounds interpretation against region-specific clinical guidelines vi…

---

## [EHRQA from MedGemma](https://discuss.ai.google.dev/t/ehrqa-from-medgemma/96605)

<div class="topic-metadata">

**Author:** [@shivak](https://discuss.ai.google.dev/u/shivak)\
**Replies:** 5\
**Last updated:** [August 20, 2026, 3:31pm UTC](https://discuss.ai.google.dev/t/ehrqa-from-medgemma/96605 "2026-08-20T15:31:13Z")

</div>

Are there plans to release the EHRQA (Synthea) benchmark which was created and used in the MedGemma tech report? I appreciate your recent research contributions here.

---

## [MEDASR for other languages?](https://discuss.ai.google.dev/t/medasr-for-other-languages/141397)

<div class="topic-metadata">

**Author:** [@Alexcc](https://discuss.ai.google.dev/u/Alexcc)\
**Replies:** 4\
**Last updated:** [August 12, 2026, 4:49pm UTC](https://discuss.ai.google.dev/t/medasr-for-other-languages/141397 "2026-08-12T16:49:33Z")

</div>

Congrats to the HAI-DEF team for releasing MedASR which is a nice open source model competing with the well known Dragon Medical One product. I have read that you plan to release eventually a multi-language support/chec…

---

## [Integrating MedGemma with LangChain Challenges and Solutions](https://discuss.ai.google.dev/t/integrating-medgemma-with-langchain-challenges-and-solutions/96247)

<div class="topic-metadata">

**Author:** [@zeerakwyne](https://discuss.ai.google.dev/u/zeerakwyne)\
**Replies:** 4\
**Last updated:** [August 10, 2026, 6:52am UTC](https://discuss.ai.google.dev/t/integrating-medgemma-with-langchain-challenges-and-solutions/96247 "2026-08-10T06:52:24Z")

</div>

Hi everyone, this is Zeerak from Spryt. We’re working on a multi-agent healthcare system for appointment scheduling and patient support, built using LangChain and LangGraph. Our system uses a sophisticated architecture …

---

## [MedGemma 27b text model to Vertex AI Endpoint Deployment Failure – “Model server exited unexpectedly”](https://discuss.ai.google.dev/t/medgemma-27b-text-model-to-vertex-ai-endpoint-deployment-failure-model-server-exited-unexpectedly/111504)

<div class="topic-metadata">

**Author:** [@Ankit\_Mishra](https://discuss.ai.google.dev/u/Ankit_Mishra)\
**Replies:** 2\
**Last updated:** [August 3, 2026, 11:37am UTC](https://discuss.ai.google.dev/t/medgemma-27b-text-model-to-vertex-ai-endpoint-deployment-failure-model-server-exited-unexpectedly/111504 "2026-08-03T11:37:49Z")

</div>

Description: I have a MedGemma 27b text fine-tuned model packaged as a .mar file (TorchServe format). The .mar file is uploaded to Google Cloud Storage and successfully imported into Vertex AI Model Registry. When I at…

---

## [RAG pipeline over large clinical PDFs](https://discuss.ai.google.dev/t/rag-pipeline-over-large-clinical-pdfs/172518)

<div class="topic-metadata">

**Author:** [@HH1](https://discuss.ai.google.dev/u/HH1)\
**Replies:** 3\
**Last updated:** [July 15, 2026, 6:45pm UTC](https://discuss.ai.google.dev/t/rag-pipeline-over-large-clinical-pdfs/172518 "2026-07-15T18:45:07Z")

</div>

Context: We are building a tool that retrieves clinical protocols from large (500+ page) PDFs and shows clinicians relevant protocols in real time. Questions: What splitting/parsing methods are best for large clinic…

---

## [Question on pre-training for medical tasks](https://discuss.ai.google.dev/t/question-on-pre-training-for-medical-tasks/109271)

<div class="topic-metadata">

**Author:** [@NicoleMcNally](https://discuss.ai.google.dev/u/NicoleMcNally)\
**Replies:** 4\
**Last updated:** [July 1, 2026, 12:06am UTC](https://discuss.ai.google.dev/t/question-on-pre-training-for-medical-tasks/109271 "2026-07-01T00:06:05Z")

</div>

In your technical report you articulate that mixing general-purpose data during the ‘Vision Encoder Enhancement’ and ‘Multimodal Decoder Pretraining’ stages was crucial ‘to preserve the visual-language reasoning capabili…

---

## [Has there been any other News Letters from Google AI for Developers (HAI-DEF) after January 2026)? I didn't receive any since then](https://discuss.ai.google.dev/t/has-there-been-any-other-news-letters-from-google-ai-for-developers-hai-def-after-january-2026-i-didnt-receive-any-since-then/167251)

<div class="topic-metadata">

**Author:** [@i\_NewTechs](https://discuss.ai.google.dev/u/i_NewTechs)\
**Replies:** 3\
**Last updated:** [June 17, 2026, 12:48pm UTC](https://discuss.ai.google.dev/t/has-there-been-any-other-news-letters-from-google-ai-for-developers-hai-def-after-january-2026-i-didnt-receive-any-since-then/167251 "2026-06-17T12:48:24Z")

</div>

I didn’t receive any News Letters from Google AI for Developers (HAI-DEF) after January 2026)? Is there any News Letters sent out since then?

---

## [MedASR – Outstanding Feedback from Clinicians](https://discuss.ai.google.dev/t/medasr-outstanding-feedback-from-clinicians/169929)

<div class="topic-metadata">

**Author:** [@Mary\_Colm](https://discuss.ai.google.dev/u/Mary_Colm)\
**Replies:** 4\
**Last updated:** [June 10, 2026, 12:40pm UTC](https://discuss.ai.google.dev/t/medasr-outstanding-feedback-from-clinicians/169929 "2026-06-10T12:40:46Z")

</div>

Hello Google HAI-DEF team, Thank you for releasing MedASR , it’s a fantastic model and has been very valuable for our work. We recently built and released an iOS offline-first AI scribe called Medavo, which uses MedASR…

---

## [Launch of HAI-DEF Showcase - a space dedicated to your work based on HAI-DEF open assets](https://discuss.ai.google.dev/t/launch-of-hai-def-showcase-a-space-dedicated-to-your-work-based-on-hai-def-open-assets/170034)

<div class="topic-metadata">

**Author:** [@fmahvar](https://discuss.ai.google.dev/u/fmahvar)\
**Replies:** 0\
**Last updated:** [June 7, 2026, 8:05pm UTC](https://discuss.ai.google.dev/t/launch-of-hai-def-showcase-a-space-dedicated-to-your-work-based-on-hai-def-open-assets/170034 "2026-06-07T20:05:43Z")

</div>

We are thrilled to announce the launch of the HAI-DEF Showcase, a dedicated space to celebrate the incredible projects you’ve been creating with our open models! We have a lineup currently in the publishing pipeline, and…

---

## [MedGemma Impact Challenge participant about to lose GCP account over Vertex AI billing dispute](https://discuss.ai.google.dev/t/medgemma-impact-challenge-participant-about-to-lose-gcp-account-over-vertex-ai-billing-dispute/123627)

<div class="topic-metadata">

**Author:** [@DrJCF](https://discuss.ai.google.dev/u/DrJCF)\
**Replies:** 4\
**Last updated:** [May 18, 2026, 12:30pm UTC](https://discuss.ai.google.dev/t/medgemma-impact-challenge-participant-about-to-lose-gcp-account-over-vertex-ai-billing-dispute/123627 "2026-05-18T12:30:55Z")

</div>

I’m a practicing surgeon (25 years), AI2030 Global Fellow, Co-Editor-in-Chief at StatPearls, and CEO of EdAI Systems. I’m participating in the MedGemma Impact Challenge building PMAI-Gemma, a privacy-preserving personal …

---

## [Lost in Translation: Building Reliable Clinical AI Pipelines](https://discuss.ai.google.dev/t/lost-in-translation-building-reliable-clinical-ai-pipelines/142952)

<div class="topic-metadata">

**Author:** [@Navin\_Kumar1](https://discuss.ai.google.dev/u/Navin_Kumar1)\
**Replies:** 1\
**Last updated:** [May 18, 2026, 11:37am UTC](https://discuss.ai.google.dev/t/lost-in-translation-building-reliable-clinical-ai-pipelines/142952 "2026-05-18T11:37:42Z")

</div>

We recently explored how ambiguity in clinical requests impacts downstream AI workflows and how we can address it with structured systems. In practice, many workflows begin with loosely defined requests (e.g., “predict …

---

## [Supporting Medical Imaging AI with High-Quality Annotation & Segmentation (Free Pilot Available)](https://discuss.ai.google.dev/t/supporting-medical-imaging-ai-with-high-quality-annotation-segmentation-free-pilot-available/137797)

<div class="topic-metadata">

**Author:** [@Safna\_Kallingal](https://discuss.ai.google.dev/u/Safna_Kallingal)\
**Replies:** 1\
**Last updated:** [May 15, 2026, 9:48am UTC](https://discuss.ai.google.dev/t/supporting-medical-imaging-ai-with-high-quality-annotation-segmentation-free-pilot-available/137797 "2026-05-15T09:48:27Z")

</div>

Hi everyone,:waving\_hand: I’ve been following the progress of Health AI Developer Foundations and MedGemma- exciting to see how rapidly real-world medical imaging applications are evolving. I’m part of the team at Medr…

---

## [When do you anticipate an updated model for MedGemma?](https://discuss.ai.google.dev/t/when-do-you-anticipate-an-updated-model-for-medgemma/143159)

<div class="topic-metadata">

**Author:** [@bob\_sacamano](https://discuss.ai.google.dev/u/bob_sacamano)\
**Replies:** 1\
**Last updated:** [May 15, 2026, 9:44am UTC](https://discuss.ai.google.dev/t/when-do-you-anticipate-an-updated-model-for-medgemma/143159 "2026-05-15T09:44:38Z")

</div>

I am using this for radiology and curious when an updated model will come out. I was impressed by the last model

---

## [MedASR: Clarification Needed on Handling of Brace Tokens and Preprocessing Rules for Fine-Tuning & Decoding](https://discuss.ai.google.dev/t/medasr-clarification-needed-on-handling-of-brace-tokens-and-preprocessing-rules-for-fine-tuning-decoding/116107)

<div class="topic-metadata">

**Author:** [@csetanmayjain](https://discuss.ai.google.dev/u/csetanmayjain)\
**Replies:** 3\
**Last updated:** [April 29, 2026, 7:12pm UTC](https://discuss.ai.google.dev/t/medasr-clarification-needed-on-handling-of-brace-tokens-and-preprocessing-rules-for-fine-tuning-decoding/116107 "2026-04-29T19:12:59Z")

</div>

Hi Team, I am currently benchmarking and experimenting with the google/medasr model using the provided evaluation pipeline. During this process, I noticed that many transcripts in the medical ASR domain contain special …

---

## [MedGemma for Clinical AI Model Selection and Deployment](https://discuss.ai.google.dev/t/medgemma-for-clinical-ai-model-selection-and-deployment/134969)

<div class="topic-metadata">

**Author:** [@Navin\_Kumar1](https://discuss.ai.google.dev/u/Navin_Kumar1)\
**Replies:** 2\
**Last updated:** [March 26, 2026, 12:48pm UTC](https://discuss.ai.google.dev/t/medgemma-for-clinical-ai-model-selection-and-deployment/134969 "2026-03-26T12:48:55Z")

</div>

We recently explored using MedGemma within a multi-agent clinical AI system to help with both model selection and deployment. In most workflows today, picking the right model is still a very manual process, and deployme…

---

## [MedGemma 4B: Safety guardrails, prompting strategies, and on-device performance for a clinical triage app](https://discuss.ai.google.dev/t/medgemma-4b-safety-guardrails-prompting-strategies-and-on-device-performance-for-a-clinical-triage-app/134337)

<div class="topic-metadata">

**Author:** [@Bronson\_Bakunga](https://discuss.ai.google.dev/u/Bronson_Bakunga)\
**Replies:** 2\
**Last updated:** [March 23, 2026, 10:03am UTC](https://discuss.ai.google.dev/t/medgemma-4b-safety-guardrails-prompting-strategies-and-on-device-performance-for-a-clinical-triage-app/134337 "2026-03-23T10:03:37Z")

</div>

Hello everyone, We’re building an on-device AI triage tool for frontline health workers in Uganda. The app runs entirely on the user’s Android device using MedGemma 4B. We’ve taken time to make sure the questions below …

---

## [Missing the submission deadline for MedGemma Impact Challenge due to snowstorm](https://discuss.ai.google.dev/t/missing-the-submission-deadline-for-medgemma-impact-challenge-due-to-snowstorm/126728)

<div class="topic-metadata">

**Author:** [@Oluwaferanmi\_Oyelude](https://discuss.ai.google.dev/u/Oluwaferanmi_Oyelude)\
**Replies:** 1\
**Last updated:** [February 26, 2026, 4:53am UTC](https://discuss.ai.google.dev/t/missing-the-submission-deadline-for-medgemma-impact-challenge-due-to-snowstorm/126728 "2026-02-26T04:53:04Z")

</div>

Hi Organizing Team, I am writing to see if there is any possibility of a late submission for my project, Ward.AI, which I completed for the MedGemma Impact Challenge. My submission was delayed to an emergency travel sit…

---

## [MedGemma for Improving Chronic Pain Outcomes](https://discuss.ai.google.dev/t/medgemma-for-improving-chronic-pain-outcomes/123665)

<div class="topic-metadata">

**Author:** [@Navin\_Kumar1](https://discuss.ai.google.dev/u/Navin_Kumar1)\
**Replies:** 2\
**Last updated:** [February 20, 2026, 3:26pm UTC](https://discuss.ai.google.dev/t/medgemma-for-improving-chronic-pain-outcomes/123665 "2026-02-20T15:26:14Z")

</div>

We recently used MedGemma within a multi-agent system, for improving chronic pain and opioid use disorder outcomes. MedGemma was tasked with assessing whether a clinically defined pain spike occurred relative to patient…

---

## [MEDASR: Audio Length](https://discuss.ai.google.dev/t/medasr-audio-length/122940)

<div class="topic-metadata">

**Author:** [@kamalkraj](https://discuss.ai.google.dev/u/kamalkraj)\
**Replies:** 1\
**Last updated:** [February 19, 2026, 10:53am UTC](https://discuss.ai.google.dev/t/medasr-audio-length/122940 "2026-02-19T10:53:42Z")

</div>

What is the recommended size of MEDASR audio length for transcription. Or For all the audio is it recommended to use 20s sliding window with 2s overlap ?

---

## [Detecting PHI from DICOM or PNG images](https://discuss.ai.google.dev/t/detecting-phi-from-dicom-or-png-images/118239)

<div class="topic-metadata">

**Author:** [@Mehul\_Sampat](https://discuss.ai.google.dev/u/Mehul_Sampat)\
**Replies:** 2\
**Last updated:** [January 24, 2026, 10:27pm UTC](https://discuss.ai.google.dev/t/detecting-phi-from-dicom-or-png-images/118239 "2026-01-24T22:27:30Z")

</div>

Hi All! I really enjoy working with the MedGemma 1.5 release. The github notebooks are awesome. In one notebook we see it can detect the left clavicle and also put a box around it. I have an open-source DICOM with pse…

---

## [Introducing the Next generation medical image interpretation with MedGemma 1.5](https://discuss.ai.google.dev/t/introducing-the-next-generation-medical-image-interpretation-with-medgemma-1-5/115538)

<div class="topic-metadata">

**Author:** [@pradeepkuppala](https://discuss.ai.google.dev/u/pradeepkuppala)\
**Replies:** 0\
**Last updated:** [January 15, 2026, 8:51pm UTC](https://discuss.ai.google.dev/t/introducing-the-next-generation-medical-image-interpretation-with-medgemma-1-5/115538 "2026-01-15T20:51:38Z")

</div>

Next generation medical image interpretation with MedGemma 1.5 and medical speech to text with MedASR. We are updating our open MedGemma model with improved medical imaging support. We also describe MedASR, our new open …

---

## [Provenance of KenLM model included with MedASR](https://discuss.ai.google.dev/t/provenance-of-kenlm-model-included-with-medasr/113541)

<div class="topic-metadata">

**Author:** [@Ryan\_Walklin](https://discuss.ai.google.dev/u/Ryan_Walklin)\
**Replies:** 3\
**Last updated:** [January 9, 2026, 11:37pm UTC](https://discuss.ai.google.dev/t/provenance-of-kenlm-model-included-with-medasr/113541 "2026-01-09T23:37:02Z")

</div>

Hi, Thanks for your work on MedASR, I’ve done some preliminary work with it and it looks very promising. I’m curious about the source/training data for the KenLM model provided for beam search, can you share any details…

[Next page](https://discuss.ai.google.dev/c/hai-def/62.md?page=1)
