{"id":2864,"date":"2025-12-10T09:55:17","date_gmt":"2025-12-10T09:55:17","guid":{"rendered":"https:\/\/dr7.ai\/blog\/?p=2864"},"modified":"2025-12-22T09:51:55","modified_gmt":"2025-12-22T09:51:55","slug":"google-medgemma-practical-guide-for-ai-engineers","status":"publish","type":"post","link":"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/","title":{"rendered":"Google MedGemma: Practical Guide for AI Engineers"},"content":{"rendered":"\n<p>When I first pulled MedGemma 4B into a clinical sandbox, my goal was simple: see if an open-weight model could survive the same red\u2011team prompts I usually reserve for commercial, closed LLMs. Within a weekend, I had it summarizing de\u2011identified oncology notes, critiquing draft radiology reports, and, just as important, clearly exposing where it still hallucinated.<\/p>\n\n\n\n<p>In this <a href=\"https:\/\/developers.google.com\/health-ai-developer-foundations\/medgemma\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">MedGemma tutorial<\/a>, I&#8217;ll walk through how I think about the model family as a <a href=\"https:\/\/developers.google.com\/health-ai-developer-foundations\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">health\u2011AI engineer<\/a>: architectures, benchmarks, GPU sizing, deployment options, fine\u2011tuning strategy, and safety controls. The focus is practical: code\u2011ready guidance to help you de\u2011risk real clinical integrations under HIPAA\/GDPR, not a glossy model launch recap.<\/p>\n\n\n\n<p>Medical disclaimer: Nothing here is medical advice. MedGemma is a research\/development tool (as of late 2025) and must not be used to make or replace clinical decisions without appropriate regulatory clearance, institutional review, and supervision by qualified clinicians.<\/p>\n\n\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_76 ez-toc-wrap-left counter-hierarchy ez-toc-counter ez-toc-transparent ez-toc-container-direction\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">Table of Contents<\/p>\n<label for=\"ez-toc-cssicon-toggle-item-69e1d126b5e61\" class=\"ez-toc-cssicon-toggle-label\"><span class=\"ez-toc-cssicon\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/label><input type=\"checkbox\"  id=\"ez-toc-cssicon-toggle-item-69e1d126b5e61\"  aria-label=\"Toggle\" \/><nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Understanding_MedGemma_Googles_New_Standard_for_Open_Medical_AI\" >Understanding MedGemma: Google&#8217;s New Standard for Open Medical AI<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#From_Med-PaLM_to_MedGemma_The_Evolution_of_Medical_LLMs\" >From Med-PaLM to MedGemma: The Evolution of Medical LLMs<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Open_Weights_vs_Closed_API_Why_MedGemma_Matters\" >Open Weights vs. Closed API: Why MedGemma Matters<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Deep_Dive_MedGemma_Architectures_Benchmarks\" >Deep Dive: MedGemma Architectures &amp; Benchmarks<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#MedGemma_4B_vs_27B_Choosing_Between_Speed_and_Reasoning\" >MedGemma 4B vs 27B: Choosing Between Speed and Reasoning<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Performance_Analysis_MedQA_Scores_and_Clinical_Accuracy\" >Performance Analysis: MedQA Scores and Clinical Accuracy<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Inside_the_Tech_Multimodal_Capabilities\" >Inside the Tech: Multimodal Capabilities<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Mastering_Medical_Imagery_with_MedSigLIP\" >Mastering Medical Imagery with MedSigLIP<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Processing_EHRs_and_Clinical_Notes\" >Processing EHRs and Clinical Notes<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Developer_Guide_How_to_Run_MedGemma\" >Developer Guide: How to Run MedGemma<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Deployment_Options_Hugging_Face_Kaggle_and_Vertex_AI\" >Deployment Options: Hugging Face, Kaggle, and Vertex AI<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Environment_Setup_GPU_Requirements_for_4B_and_27B\" >Environment Setup: GPU Requirements for 4B and 27B<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-13\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Fine-Tuning_for_Specialized_Healthcare_Tasks\" >Fine-Tuning for Specialized Healthcare Tasks<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-14\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Step-by-Step_LoRA_Fine-Tuning_on_Custom_Medical_Data\" >Step-by-Step: LoRA Fine-Tuning on Custom Medical Data<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-15\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Data_Privacy_and_De-identification_Best_Practices\" >Data Privacy and De-identification Best Practices<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-16\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Real-World_Use_Cases\" >Real-World Use Cases<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-17\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Building_an_Automated_Radiology_Report_Assistant\" >Building an Automated Radiology Report Assistant<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-18\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Clinical_Decision_Support_Systems_CDSS\" >Clinical Decision Support Systems (CDSS)<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-19\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Limitations_and_Responsible_AI_Usage\" >Limitations and Responsible AI Usage<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-20\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Handling_Hallucinations_in_Medical_Contexts\" >Handling Hallucinations in Medical Contexts<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-21\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Googles_Safety_Filters_and_Use_Restrictions\" >Google&#8217;s Safety Filters and Use Restrictions<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-22\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Conclusion_The_Future_of_Open_Source_in_Healthcare\" >Conclusion: The Future of Open Source in Healthcare<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-23\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#MedGemma_Tutorial_%E2%80%93_Frequently_Asked_Questions\" >MedGemma Tutorial \u2013 Frequently Asked Questions<\/a><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><ul class='ez-toc-list-level-4' ><li class='ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-24\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#What_is_MedGemma_and_who_is_this_MedGemma_tutorial_for\" >What is MedGemma and who is this MedGemma tutorial for?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-25\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#How_do_I_choose_between_MedGemma_4B_and_MedGemma_27B_for_my_project\" >How do I choose between MedGemma 4B and MedGemma 27B for my project?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-26\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#How_do_I_actually_run_MedGemma_step%E2%80%91by%E2%80%91step\" >How do I actually run MedGemma step\u2011by\u2011step?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-27\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#Can_I_use_MedGemma_for_real_clinical_decision_making_today\" >Can I use MedGemma for real clinical decision making today?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-4'><a class=\"ez-toc-link ez-toc-heading-28\" href=\"https:\/\/dr7.ai\/blog\/health\/google-medgemma-practical-guide-for-ai-engineers\/#What_are_best_practices_for_fine%E2%80%91tuning_MedGemma_on_my_hospitals_data\" >What are best practices for fine\u2011tuning MedGemma on my hospital\u2019s data?<\/a><\/li><\/ul><\/li><\/ul><\/li><\/ul><\/nav><\/div>\n<h2 class=\"wp-block-heading\" id=\"understanding-medgemma-googles-new-standard-for-open-medical-ai\"><span class=\"ez-toc-section\" id=\"Understanding_MedGemma_Googles_New_Standard_for_Open_Medical_AI\"><\/span>Understanding MedGemma: Google&#8217;s New Standard for Open Medical AI<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n<p>MedGemma is <a href=\"https:\/\/deepmind.google\/models\/gemma\/medgemma\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">Google Health&#8217;s open\u2011weight medical LLM line<\/a>, released in 2025 as part of the Gemma ecosystem. It&#8217;s trained on a mix of biomedical literature, medical question\u2011answer pairs, and synthetic data, and it comes in both text\u2011only and multimodal variants.<\/p>\n\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-1 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><img fetchpriority=\"high\" decoding=\"async\" width=\"1024\" height=\"511\" data-id=\"2867\" src=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-1-1-1024x511.png\" alt=\"Google DeepMind MedGemma introduction page highlighting &quot;A Gemma 3 variant optimized for medical text and image comprehension&quot;\" class=\"wp-image-2867\" srcset=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-1-1-1024x511.png 1024w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-1-1-300x150.png 300w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-1-1-768x383.png 768w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-1-1.png 1280w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/figure>\n\n\n<h3 class=\"wp-block-heading\" id=\"from-medpalm-to-medgemma-the-evolution-of-medical-llms\"><span class=\"ez-toc-section\" id=\"From_Med-PaLM_to_MedGemma_The_Evolution_of_Medical_LLMs\"><\/span>From Med-PaLM to MedGemma: The Evolution of Medical LLMs<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>I think of Med-PaLM as the &#8220;proof of concept&#8221; and MedGemma as the &#8220;developer\u2011ready&#8221; successor. Med-PaLM and Med-PaLM 2 showed that large generalist models could hit clinician\u2011level performance on MedQA and similar exams, but they were closed and hard to integrate in regulated pipelines.<\/p>\n\n\n\n<p>MedGemma keeps much of that medical reasoning focus but exposes:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Open weights (see the <a href=\"https:\/\/github.com\/Google-Health\/medgemma\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">official GitHub<\/a> and Hugging Face repos) for inspection and on\u2011prem deployment.<\/li>\n\n\n\n<li>Smaller variants (4B, 27B) that are realistically runnable on hospital or cloud GPUs.<\/li>\n\n\n\n<li>Multimodal capabilities via MedSigLIP for imaging tasks like radiographs and pathology slides.<\/li>\n<\/ul>\n\n\n<h3 class=\"wp-block-heading\" id=\"open-weights-vs-closed-api-why-medgemma-matters\"><span class=\"ez-toc-section\" id=\"Open_Weights_vs_Closed_API_Why_MedGemma_Matters\"><\/span>Open Weights vs. Closed API: Why MedGemma Matters<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>For regulated markets, open weights change the threat model:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>You can keep PHI in your own VPC or on\u2011prem cluster, avoiding cross\u2011border data transfer issues under HIPAA\/GDPR.<\/li>\n\n\n\n<li>You can instrument and log everything, tokens, prompts, responses, for validation and incident review.<\/li>\n\n\n\n<li>You can fine\u2011tune and constrain the model using LoRA or adapters, rather than praying a vendor updates their safety filters.<\/li>\n<\/ul>\n\n\n\n<p>The tradeoff is responsibility: you now own patching, access control, and safety evaluation. That&#8217;s exactly why a disciplined evaluation and deployment process is non\u2011negotiable with MedGemma.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"deep-dive-medgemma-architectures-amp-benchmarks\"><span class=\"ez-toc-section\" id=\"Deep_Dive_MedGemma_Architectures_Benchmarks\"><\/span>Deep Dive: MedGemma Architectures &amp; Benchmarks<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h3 class=\"wp-block-heading\" id=\"medgemma-4b-vs-27b-choosing-between-speed-and-reasoning\"><span class=\"ez-toc-section\" id=\"MedGemma_4B_vs_27B_Choosing_Between_Speed_and_Reasoning\"><\/span>MedGemma 4B vs 27B: Choosing Between Speed and Reasoning<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>MedGemma currently ships in two main instruction\u2011tuned text variants:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/huggingface.co\/google\/medgemma-4b-it\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">medgemma-4b-it<\/a> \u2013 ~4B parameters, efficient, good for latency\u2011sensitive pipelines, edge experimentation, or as a reranker\/assistant around a larger model.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-2 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1024\" height=\"510\" data-id=\"2868\" src=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-2-1-1024x510.png\" alt=\"Hugging Face repository for google\/medgemma-4b-it, 4B multimodal medical model based on Gemma 3, requiring Health AI terms acceptance\" class=\"wp-image-2868\" srcset=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-2-1-1024x510.png 1024w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-2-1-300x149.png 300w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-2-1-768x382.png 768w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-2-1.png 1280w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/huggingface.co\/google\/medgemma-27b-text-it\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">medgemma-27b-text-it<\/a> \u2013 ~27B parameters, noticeably stronger on long\u2011form reasoning, edge cases, and exam\u2011style questions.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-3 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1024\" height=\"502\" data-id=\"2871\" src=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-3-1-1024x502.png\" alt=\"Hugging Face google\/medgemma-27b-text-it page, 27B parameter text-only MedGemma variant derived from Gemma 3 27B\" class=\"wp-image-2871\" srcset=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-3-1-1024x502.png 1024w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-3-1-300x147.png 300w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-3-1-768x376.png 768w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-3-1.png 1280w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/figure>\n\n\n\n<p>In my testing on de\u2011identified endocrine consult notes, 4B worked for classification and summarization, while 27B added value for differential diagnosis explanation and guideline citation. If you&#8217;re building production decision support, I&#8217;d prototype with 4B but budget GPU for 27B.<\/p>\n\n\n<h3 class=\"wp-block-heading\" id=\"performance-analysis-medqa-scores-and-clinical-accuracy\"><span class=\"ez-toc-section\" id=\"Performance_Analysis_MedQA_Scores_and_Clinical_Accuracy\"><\/span>Performance Analysis: MedQA Scores and Clinical Accuracy<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>According to <a href=\"https:\/\/arxiv.org\/abs\/2507.05201\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">Google&#8217;s 2025 MedGemma paper<\/a> and <a href=\"https:\/\/www.google.com\/url?q=https%3A%2F%2Fresearch.google%2Fblog%2Fmedgemma-our-most-capable-open-models-for-health-ai-development%2F\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">blog<\/a>, 27B matches or exceeds earlier Med-PaLM\u2011class models on:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>MedQA (USMLE\u2011style) multiple\u2011choice exams<\/li>\n\n\n\n<li>MultiMedQA\u2011style composite benchmarks<\/li>\n\n\n\n<li>Several radiology and ophthalmology tasks (for multimodal variants)<\/li>\n<\/ul>\n\n\n\n<p>Benchmarks give a ceiling: your real risk lies in distribution shift and hallucinations. I recommend:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Building a local evaluation set of de\u2011identified cases mapped to guideline\u2011backed answers.<\/li>\n\n\n\n<li>Tracking factual accuracy, citation fidelity, and refusal rates per specialty.<\/li>\n\n\n\n<li>Logging adverse outputs (unsafe, biased, off\u2011label) and feeding them back into prompt\/adapter tuning.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-4 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"551\" data-id=\"2866\" src=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/593b7a53-678a-46cc-a5c4-6181fcbd9d69-1024x551.png\" alt=\"MedQA benchmark leaderboard (2025) showing MedGemma-27B leading open-source medical LLMs in accuracy vs cost per million tokens\" class=\"wp-image-2866\" srcset=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/593b7a53-678a-46cc-a5c4-6181fcbd9d69-1024x551.png 1024w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/593b7a53-678a-46cc-a5c4-6181fcbd9d69-300x161.png 300w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/593b7a53-678a-46cc-a5c4-6181fcbd9d69-768x413.png 768w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/593b7a53-678a-46cc-a5c4-6181fcbd9d69.png 1250w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/figure>\n\n\n<h2 class=\"wp-block-heading\" id=\"inside-the-tech-multimodal-capabilities\"><span class=\"ez-toc-section\" id=\"Inside_the_Tech_Multimodal_Capabilities\"><\/span>Inside the Tech: Multimodal Capabilities<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h3 class=\"wp-block-heading\" id=\"mastering-medical-imagery-with-medsiglip\"><span class=\"ez-toc-section\" id=\"Mastering_Medical_Imagery_with_MedSigLIP\"><\/span>Mastering Medical Imagery with MedSigLIP<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>MedGemma&#8217;s vision stack leverages MedSigLIP, a medical\u2011adapted variant of Google&#8217;s SigLIP. In practice, you pass an image embedding into the language model, then condition generation on both modalities.<\/p>\n\n\n\n<p>In a radiology sandbox, I used the multimodal MedGemma to:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Describe key findings on de\u2011identified chest X\u2011rays.<\/li>\n\n\n\n<li>Compare current vs prior images (e.g., nodule interval growth).<\/li>\n\n\n\n<li>Generate draft impressions for an attending to edit.<\/li>\n<\/ul>\n\n\n\n<p>Two caveats:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li>As of late 2025, these models are not FDA\u2011cleared devices: they&#8217;re research tools.<\/li>\n\n\n\n<li>I always required the model to output: &#8220;This is not a diagnostic report. A radiologist must review the images.&#8221;<\/li>\n<\/ol>\n\n\n<h3 class=\"wp-block-heading\" id=\"processing-ehrs-and-clinical-notes\"><span class=\"ez-toc-section\" id=\"Processing_EHRs_and_Clinical_Notes\"><\/span>Processing EHRs and Clinical Notes<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>Text\u2011only MedGemma works well on structured + free\u2011text EHR exports:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Creating sectioned summaries (HPI, meds, labs, assessment\/plan).<\/li>\n\n\n\n<li>Normalizing problems to SNOMED\/ICD\u201110 with a terminology layer on top.<\/li>\n\n\n\n<li>Flagging missing guideline\u2011recommended labs.<\/li>\n<\/ul>\n\n\n\n<p>For PHI, I strongly prefer a pattern of: de\u2011identify \u2192 process with MedGemma \u2192 re\u2011link via pseudonyms inside your secure environment.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"developer-guide-how-to-run-medgemma\"><span class=\"ez-toc-section\" id=\"Developer_Guide_How_to_Run_MedGemma\"><\/span>Developer Guide: How to Run MedGemma<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h3 class=\"wp-block-heading\" id=\"deployment-options-hugging-face-kaggle-and-vertex-ai\"><span class=\"ez-toc-section\" id=\"Deployment_Options_Hugging_Face_Kaggle_and_Vertex_AI\"><\/span>Deployment Options: Hugging Face, Kaggle, and Vertex AI<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>You can get started three main ways:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Hugging Face \u2013 Pull <code>google\/medgemma-4b-it<\/code> or <code>google\/medgemma-27b-text-it<\/code> via <code>transformers<\/code> or <code>vLLM<\/code>. Good for on\u2011prem and VPC deployments.<\/li>\n\n\n\n<li>Kaggle notebooks \u2013 Google&#8217;s example notebooks are handy for quick, free GPU prototyping (no PHI, obviously).<\/li>\n\n\n\n<li>Vertex AI \u2013 For GCP shops, Gemma\/MedGemma integration lets you deploy managed endpoints, add IAM, and wrap with Cloud Logging and Policy Intelligence.<\/li>\n<\/ul>\n\n\n\n<p>In my own hospital PoC, I used Hugging Face + vLLM on Kubernetes inside a private subnet and proxied requests through an audited API gateway.<\/p>\n\n\n\n<p>Quick alternative for prototyping: Instant access to MedGemma via <a href=\"https:\/\/dr7.ai\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">dr7.ai<\/a>&#8216;s unified API (free tier, no card needed).<\/p>\n\n\n<h3 class=\"wp-block-heading\" id=\"environment-setup-gpu-requirements-for-4b-and-27b\"><span class=\"ez-toc-section\" id=\"Environment_Setup_GPU_Requirements_for_4B_and_27B\"><\/span>Environment Setup: GPU Requirements for 4B and 27B<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>Very roughly (check official docs for updates):<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>4B: fits on a single 24 GB GPU (A10, L4) with 4\u2011bit quantization: great for dev and small batch inference.<\/li>\n\n\n\n<li>27B: more comfortable on 2\u00d724 GB or a single 80 GB GPU (A100\/H100), especially for long sequences.<\/li>\n<\/ul>\n\n\n\n<p>I recommend:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Using vLLM or TensorRT\u2011LLM for production inference.<\/li>\n\n\n\n<li>Pinning container images with explicit CUDA\/cuDNN versions.<\/li>\n\n\n\n<li>Adding resource limits and per\u2011tenant rate limiting to avoid denial\u2011of\u2011service and unexpected latency spikes.<\/li>\n<\/ul>\n\n\n<h2 class=\"wp-block-heading\" id=\"finetuning-for-specialized-healthcare-tasks\"><span class=\"ez-toc-section\" id=\"Fine-Tuning_for_Specialized_Healthcare_Tasks\"><\/span>Fine-Tuning for Specialized Healthcare Tasks<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h3 class=\"wp-block-heading\" id=\"stepbystep-lora-finetuning-on-custom-medical-data\"><span class=\"ez-toc-section\" id=\"Step-by-Step_LoRA_Fine-Tuning_on_Custom_Medical_Data\"><\/span>Step-by-Step: LoRA Fine-Tuning on Custom Medical Data<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>My default recipe for MedGemma specialization is parameter\u2011efficient fine\u2011tuning (PEFT) with LoRA:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li>Curate de\u2011identified data: e.g., cardiology notes paired with guideline\u2011aligned summaries.<\/li>\n\n\n\n<li>Define a narrow task: &#8220;Summarize clinic notes for communication with PCPs,&#8221; not &#8220;general cardiology expert.&#8221;<\/li>\n\n\n\n<li>Use <code>peft<\/code> + <code>transformers<\/code> to attach LoRA adapters to attention and MLP layers.<\/li>\n\n\n\n<li>Train on instruction\u2013response pairs with strong system prompts enforcing scope and disclaimers.<\/li>\n\n\n\n<li>Evaluate against a held\u2011out, clinician\u2011reviewed set before touching any real workflow.<\/li>\n<\/ol>\n\n\n\n<p>I&#8217;ve seen LoRA adapters under 1\u20132% of base parameters meaningfully improve style, terminology, and alignment to local practice.<\/p>\n\n\n<h3 class=\"wp-block-heading\" id=\"data-privacy-and-deidentification-best-practices\"><span class=\"ez-toc-section\" id=\"Data_Privacy_and_De-identification_Best_Practices\"><\/span>Data Privacy and De-identification Best Practices<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>For HIPAA\/GDPR, I treat MedGemma as untrusted with raw PHI by default:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Run a dedicated de\u2011identification pipeline (rule\u2011based + ML) before training or inference.<\/li>\n\n\n\n<li>Maintain data processing agreements and a RoPA (Record of Processing Activities).<\/li>\n\n\n\n<li>Never send rare disease case details to public endpoints or shared notebooks.<\/li>\n<\/ul>\n\n\n\n<p>If you must work with limited PHI (e.g., on\u2011prem fine\u2011tuning), involve your IRB or data protection officer, document your safeguards, and apply strict access controls and audit logging.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"realworld-use-cases\"><span class=\"ez-toc-section\" id=\"Real-World_Use_Cases\"><\/span>Real-World Use Cases<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n<figure class=\"wp-block-gallery has-nested-images columns-default is-cropped wp-block-gallery-5 is-layout-flex wp-block-gallery-is-layout-flex\">\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"927\" data-id=\"2865\" src=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/2b5ad8db-c8db-4ba3-964a-69de2e4ca3df-1024x927.png\" alt=\"MedGemma model family overview: MedSigLIP (0.4B image), MedGemma 4B multimodal, and 27B text\/multimodal variants with key use cases\" class=\"wp-image-2865\" srcset=\"https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/2b5ad8db-c8db-4ba3-964a-69de2e4ca3df-1024x927.png 1024w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/2b5ad8db-c8db-4ba3-964a-69de2e4ca3df-300x272.png 300w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/2b5ad8db-c8db-4ba3-964a-69de2e4ca3df-768x696.png 768w, https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/2b5ad8db-c8db-4ba3-964a-69de2e4ca3df.png 1250w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/figure>\n\n\n<h3 class=\"wp-block-heading\" id=\"building-an-automated-radiology-report-assistant\"><span class=\"ez-toc-section\" id=\"Building_an_Automated_Radiology_Report_Assistant\"><\/span>Building an Automated Radiology Report Assistant<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>In one pilot, I connected multimodal MedGemma to a PACS sandbox with de\u2011identified chest X\u2011rays:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>The system ingested the image + brief clinical indication.<\/li>\n\n\n\n<li>MedGemma generated a structured draft (Findings, Impression, Recommendations).<\/li>\n\n\n\n<li>An attending radiologist edited and signed off.<\/li>\n<\/ul>\n\n\n\n<p>The benefits were mostly consistency and speed for routine studies. Risks included occasional over\u2011confident descriptions of subtle findings. We mitigated this by forcing explicit uncertainty language and requiring radiologist sign\u2011off for every report. No direct patient\u2011facing outputs.<\/p>\n\n\n<h3 class=\"wp-block-heading\" id=\"clinical-decision-support-systems-cdss\"><span class=\"ez-toc-section\" id=\"Clinical_Decision_Support_Systems_CDSS\"><\/span>Clinical Decision Support Systems (CDSS)<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>For CDSS, I strongly prefer a retrieval\u2011augmented design:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Retrieve relevant guidelines (e.g., ACC\/AHA, NCCN) and local order sets.<\/li>\n\n\n\n<li>Let MedGemma summarize and contextualize those documents to the specific case.<\/li>\n\n\n\n<li>Log model rationales and expose them to clinicians for verification.<\/li>\n<\/ul>\n\n\n\n<p>Any CDSS that influences orders or diagnoses will likely fall under medical device regulations (FDA, MDR), so you&#8217;ll need formal validation and post\u2011market surveillance, not just offline benchmarking.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"limitations-and-responsible-ai-usage\"><span class=\"ez-toc-section\" id=\"Limitations_and_Responsible_AI_Usage\"><\/span>Limitations and Responsible AI Usage<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h3 class=\"wp-block-heading\" id=\"handling-hallucinations-in-medical-contexts\"><span class=\"ez-toc-section\" id=\"Handling_Hallucinations_in_Medical_Contexts\"><\/span>Handling Hallucinations in Medical Contexts<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>Even with strong MedQA performance, MedGemma will hallucinate:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Fabricated citations or study names.<\/li>\n\n\n\n<li>Confident but wrong dosing or contraindications.<\/li>\n\n\n\n<li>Over\u2011generalization from adult to pediatric patients.<\/li>\n<\/ul>\n\n\n\n<p>My mitigation toolkit:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Constrain scope (education, summarization, documentation support, not autonomous diagnosis).<\/li>\n\n\n\n<li>Use RAG with authoritative sources (FDA labels, UpToDate\u2011style content, local protocols) and require the model to quote them.<\/li>\n\n\n\n<li>Add a second pass checker model or rules engine for high\u2011risk content (meds, labs, procedures).<\/li>\n\n\n\n<li>Make it impossible for the system to act without human confirmation in any safety\u2011critical path.<\/li>\n<\/ul>\n\n\n<h3 class=\"wp-block-heading\" id=\"googles-safety-filters-and-use-restrictions\"><span class=\"ez-toc-section\" id=\"Googles_Safety_Filters_and_Use_Restrictions\"><\/span>Google&#8217;s Safety Filters and Use Restrictions<span class=\"ez-toc-section-end\"><\/span><\/h3>\n\n\n<p>Google&#8217;s documentation and licenses impose restrictions on:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Certain high\u2011risk medical uses, especially autonomous diagnosis or treatment recommendations.<\/li>\n\n\n\n<li>Re\u2011identification attempts or misuse of PHI.<\/li>\n<\/ul>\n\n\n\n<p>I treat Google&#8217;s guidance as a floor, not a ceiling. Internally, I overlay:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>A model governance process (model cards, risk assessments, approval gates).<\/li>\n\n\n\n<li>Regular safety audits with clinicians and security teams.<\/li>\n<\/ul>\n\n\n\n<p>If a patient is acutely unwell, the correct instruction to any user of your system should be: &#8220;Call emergency services or contact your clinician immediately,&#8221; not &#8220;Ask the model.&#8221;<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"conclusion-the-future-of-open-source-in-healthcare\"><span class=\"ez-toc-section\" id=\"Conclusion_The_Future_of_Open_Source_in_Healthcare\"><\/span>Conclusion: The Future of Open Source in Healthcare<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n<p>MedGemma doesn&#8217;t magically solve clinical AI, but it finally gives us serious, open medical LLMs we can inspect, stress\u2011test, and deploy under our own compliance regimes.<\/p>\n\n\n\n<p>My own pattern today is:<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li>Start with MedGemma 4B for low\u2011risk summarization and experimentation.<\/li>\n\n\n\n<li>Graduate to 27B for reasoning\u2011heavy tasks, always behind de\u2011identification and RAG.<\/li>\n\n\n\n<li>Layer LoRA adapters, governance, and strict human\u2011in\u2011the\u2011loop review for anything touching care.<\/li>\n<\/ol>\n\n\n\n<p>Used thoughtfully, MedGemma can be one of those tools. Used recklessly, it&#8217;s just another way to hallucinate with higher confidence. The difference is in how you test, constrain, and govern it, and that part is entirely on us.<\/p>\n\n\n\n<p><strong>Disclaimer:<\/strong><\/p>\n\n\n\n<p>The content on this website is for <strong>informational and educational purposes only<\/strong> and is intended to help readers understand AI technologies used in healthcare settings. It <strong>does not provide medical advice, diagnosis, treatment, or clinical guidance<\/strong>. Any medical decisions must be made by qualified healthcare professionals. AI models, tools, or workflows described here are <strong>assistive technologies<\/strong>, not substitutes for professional medical judgment. Deployment of any AI system in real clinical environments requires <strong>institutional approval, regulatory and legal review, data privacy compliance (e.g., HIPAA\/GDPR), and oversight by licensed medical personnel<\/strong>. DR7.ai and its authors assume no responsibility for actions taken based on this content.<\/p>\n\n\n<h2 class=\"wp-block-heading\" id=\"medgemma-tutorial-frequently-asked-questions\"><span class=\"ez-toc-section\" id=\"MedGemma_Tutorial_%E2%80%93_Frequently_Asked_Questions\"><\/span>MedGemma Tutorial \u2013 Frequently Asked Questions<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n<h4 class=\"wp-block-heading\" id=\"what-is-medgemma-and-who-is-this-medgemma-tutorial-for\"><span class=\"ez-toc-section\" id=\"What_is_MedGemma_and_who_is_this_MedGemma_tutorial_for\"><\/span>What is MedGemma and who is this MedGemma tutorial for?<span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n<p>MedGemma is Google Health\u2019s open\u2011weight medical large language model family, available in text\u2011only and multimodal variants. This MedGemma tutorial is aimed at health\u2011AI engineers, clinical informaticians, and technical teams who want practical, code\u2011level guidance for deploying MedGemma safely in HIPAA\/GDPR\u2011constrained environments.<\/p>\n\n\n<h4 class=\"wp-block-heading\" id=\"how-do-i-choose-between-medgemma-4b-and-medgemma-27b-for-my-project\"><span class=\"ez-toc-section\" id=\"How_do_I_choose_between_MedGemma_4B_and_MedGemma_27B_for_my_project\"><\/span>How do I choose between MedGemma 4B and MedGemma 27B for my project?<span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n<p>MedGemma 4B is smaller and faster, ideal for latency\u2011sensitive tasks like classification, summarization, or reranking. MedGemma 27B provides stronger long\u2011form reasoning, better for differential diagnoses explanations or guideline\u2011aware CDSS prototypes. A common pattern is prototyping with 4B, then allocating GPU budget for 27B in production.<\/p>\n\n\n<h4 class=\"wp-block-heading\" id=\"how-do-i-actually-run-medgemma-stepbystep\"><span class=\"ez-toc-section\" id=\"How_do_I_actually_run_MedGemma_step%E2%80%91by%E2%80%91step\"><\/span>How do I actually run MedGemma step\u2011by\u2011step?<span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n<p>You typically start by pulling the MedGemma model from Hugging Face, Kaggle, or deploying via Vertex AI. Set up GPUs (24 GB for 4B, larger or multiple GPUs for 27B), load with frameworks like transformers or vLLM, wrap it behind an API gateway, and add logging, access control, and rate limiting.<\/p>\n\n\n<h4 class=\"wp-block-heading\" id=\"can-i-use-medgemma-for-real-clinical-decision-making-today\"><span class=\"ez-toc-section\" id=\"Can_I_use_MedGemma_for_real_clinical_decision_making_today\"><\/span>Can I use MedGemma for real clinical decision making today?<span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n<p>No. MedGemma is a research and development tool, not an FDA\u2011cleared medical device. It should not replace clinician judgment or be used for autonomous diagnosis, prescribing, or triage. Safe uses today focus on summarization, documentation assistance, education, and decision support with strict human review and regulatory oversight.<\/p>\n\n\n<h4 class=\"wp-block-heading\" id=\"what-are-best-practices-for-finetuning-medgemma-on-my-hospitals-data\"><span class=\"ez-toc-section\" id=\"What_are_best_practices_for_fine%E2%80%91tuning_MedGemma_on_my_hospitals_data\"><\/span>What are best practices for fine\u2011tuning MedGemma on my hospital\u2019s data?<span class=\"ez-toc-section-end\"><\/span><\/h4>\n\n\n<p>Use parameter\u2011efficient fine\u2011tuning (e.g., LoRA) on carefully curated, de\u2011identified instruction\u2013response pairs for a narrow task. Run de\u2011identification before training, keep data on\u2011prem or in a secure VPC, involve your IRB or data protection officer, and evaluate against a clinician\u2011reviewed test set before any clinical pilot.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<p><strong>Past Review:<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-embed is-type-wp-embed is-provider-dr-7-ai-content-center wp-block-embed-dr-7-ai-content-center\"><div class=\"wp-block-embed__wrapper\">\n<blockquote class=\"wp-embedded-content\" data-secret=\"PBMAOMpyuD\"><a href=\"https:\/\/dr7.ai\/blog\/health\/abridge-vs-suki-vs-ambience-best-ai-scribe-2025\/\">Abridge vs Suki vs Ambience: Best AI Scribe 2025?<\/a><\/blockquote><iframe class=\"wp-embedded-content\" sandbox=\"allow-scripts\" security=\"restricted\" style=\"position: absolute; visibility: hidden;\" title=\"&#8220;Abridge vs Suki vs Ambience: Best AI Scribe 2025?&#8221; &#8212; Dr7.ai  Content Center\" src=\"https:\/\/dr7.ai\/blog\/health\/abridge-vs-suki-vs-ambience-best-ai-scribe-2025\/embed\/#?secret=aEAq4Um1hd#?secret=PBMAOMpyuD\" data-secret=\"PBMAOMpyuD\" width=\"500\" height=\"282\" frameborder=\"0\" marginwidth=\"0\" marginheight=\"0\" scrolling=\"no\"><\/iframe>\n<\/div><\/figure>\n\n\n\n<figure class=\"wp-block-embed is-type-wp-embed is-provider-dr-7-ai-content-center wp-block-embed-dr-7-ai-content-center\"><div class=\"wp-block-embed__wrapper\">\n<blockquote class=\"wp-embedded-content\" data-secret=\"bLMUJlESns\"><a href=\"https:\/\/dr7.ai\/blog\/health\/aws-healthscribe-vs-dax-copilot-who-wins\/\">AWS HealthScribe vs DAX Copilot: Who Wins?<\/a><\/blockquote><iframe class=\"wp-embedded-content\" sandbox=\"allow-scripts\" security=\"restricted\" style=\"position: absolute; visibility: hidden;\" title=\"&#8220;AWS HealthScribe vs DAX Copilot: Who Wins?&#8221; &#8212; Dr7.ai  Content Center\" src=\"https:\/\/dr7.ai\/blog\/health\/aws-healthscribe-vs-dax-copilot-who-wins\/embed\/#?secret=s4LMHubBu4#?secret=bLMUJlESns\" data-secret=\"bLMUJlESns\" width=\"500\" height=\"282\" frameborder=\"0\" marginwidth=\"0\" marginheight=\"0\" scrolling=\"no\"><\/iframe>\n<\/div><\/figure>\n\n\n\n<figure class=\"wp-block-embed is-type-wp-embed is-provider-dr-7-ai-content-center wp-block-embed-dr-7-ai-content-center\"><div class=\"wp-block-embed__wrapper\">\n<blockquote class=\"wp-embedded-content\" data-secret=\"9uhRKW5IFb\"><a href=\"https:\/\/dr7.ai\/blog\/health\/chatgpt-in-healthcare-safe-uses-risks-in-2025\/\">ChatGPT in Healthcare: Safe Uses &amp; Risks in 2025<\/a><\/blockquote><iframe class=\"wp-embedded-content\" sandbox=\"allow-scripts\" security=\"restricted\" style=\"position: absolute; visibility: hidden;\" title=\"&#8220;ChatGPT in Healthcare: Safe Uses &amp; Risks in 2025&#8221; &#8212; Dr7.ai  Content Center\" src=\"https:\/\/dr7.ai\/blog\/health\/chatgpt-in-healthcare-safe-uses-risks-in-2025\/embed\/#?secret=FIhCbcQ5pQ#?secret=9uhRKW5IFb\" data-secret=\"9uhRKW5IFb\" width=\"500\" height=\"282\" frameborder=\"0\" marginwidth=\"0\" marginheight=\"0\" scrolling=\"no\"><\/iframe>\n<\/div><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>When I first pulled MedGemma 4B into a clinical sandbox, my goal was simple: see if an open-weight model could survive the same red\u2011team prompts I usually reserve for commercial, closed LLMs. Within a weekend, I had it summarizing de\u2011identified oncology notes, critiquing draft radiology reports, and, just as important, clearly exposing where it still [&hellip;]<\/p>\n","protected":false},"author":4,"featured_media":2869,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_uag_custom_page_level_css":"","site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":"","beyondwords_generate_audio":"","beyondwords_project_id":"","beyondwords_content_id":"","beyondwords_preview_token":"","beyondwords_player_content":"","beyondwords_player_style":"","beyondwords_language_code":"","beyondwords_language_id":"","beyondwords_title_voice_id":"","beyondwords_body_voice_id":"","beyondwords_summary_voice_id":"","beyondwords_error_message":"","beyondwords_disabled":"","beyondwords_delete_content":"","beyondwords_podcast_id":"","beyondwords_hash":"","publish_post_to_speechkit":"","speechkit_hash":"","speechkit_generate_audio":"","speechkit_project_id":"","speechkit_podcast_id":"","speechkit_error_message":"","speechkit_disabled":"","speechkit_access_key":"","speechkit_error":"","speechkit_info":"","speechkit_response":"","speechkit_retries":"","speechkit_status":"","speechkit_updated_at":"","_speechkit_link":"","_speechkit_text":""},"categories":[7],"tags":[],"class_list":["post-2864","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-health"],"uagb_featured_image_src":{"full":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4.png",1280,701,false],"thumbnail":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4-150x150.png",150,150,true],"medium":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4-300x164.png",300,164,true],"medium_large":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4-768x421.png",768,421,true],"large":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4-1024x561.png",1024,561,true],"1536x1536":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4.png",1280,701,false],"2048x2048":["https:\/\/dr7.ai\/blog\/wp-content\/uploads\/2025\/12\/1280X1280-4.png",1280,701,false]},"uagb_author_info":{"display_name":"Andychen","author_link":"https:\/\/dr7.ai\/blog\/author\/andychen\/"},"uagb_comment_info":0,"uagb_excerpt":"When I first pulled MedGemma 4B into a clinical sandbox, my goal was simple: see if an open-weight model could survive the same red\u2011team prompts I usually reserve for commercial, closed LLMs. Within a weekend, I had it summarizing de\u2011identified oncology notes, critiquing draft radiology reports, and, just as important, clearly exposing where it still&hellip;","_links":{"self":[{"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/posts\/2864","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/comments?post=2864"}],"version-history":[{"count":2,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/posts\/2864\/revisions"}],"predecessor-version":[{"id":2933,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/posts\/2864\/revisions\/2933"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/media\/2869"}],"wp:attachment":[{"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/media?parent=2864"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/categories?post=2864"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dr7.ai\/blog\/wp-json\/wp\/v2\/tags?post=2864"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}