PaperlayerView exact revision on arXiv

Source-linked AI summary

MedGemma Technical Report

Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, Justin Chen, Fereshteh Mahvar, Liron Yatziv, Tiffany Chen, Bram Sterling, Stefanie Anna Baby, Susanna Maria Baby, Jeremy Lai, Samuel Schmidgall, Lu Yang, Kejia Chen, Per Bjornsson, Shashir Reddy, Ryan Brush, Kenneth Philbrick, Mercy Asiedu, Ines Mezerreg, Howard Hu, Howard Yang, Richa Tiwari, Sunny Jansen, Preeti Singh, Yun Liu, Shekoofeh Azizi, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Riviere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean-bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Elena Buchatskaya, Jean-Baptiste Alayrac, Dmitry Lepikhin, Vlad Feinberg, Sebastian Borgeaud, Alek Andreev, Cassidy Hardin, Robert Dadashi, Léonard Hussenot, Armand Joulin, Olivier Bachem, Yossi Matias, Katherine Chou, Avinatan Hassidim, Kavi Goel, Clement Farabet, Joelle Barral, Tris Warkentin, Jonathon Shlens, David Fleet, Victor Cotruta, Omar Sanseviero, Gus Martins, Phoebe Kirk, Anand Rao, Shravya Shetty, David F. Steiner, Can Kirmizibayrak, Rory Pilgrim, Daniel Golden, Lin Yang

arXiv:2507.05201v4cs.AIcs.CLcs.CV

TL;DR

Medical AI needs models with nuanced understanding of diverse medical data beyond generic models. MedGemma and MedSigLIP provide medically optimized vision-language capabilities that perform strongly across medical benchmarks and improve further with fine-tuning.

  • Problem

    Generic vision-language models can lack nuanced medical understanding and robust reasoning about medical data.

    Source: §1 · p. 2

  • Method

    The paper develops medically optimized MedGemma vision-language models and the MedSigLIP medical image encoder, built upon Gemma 3.

    Source: §2.2.1 · p. 4, §8 · p. 21

  • Results

    MedGemma and MedSigLIP perform strongly across diverse medical reasoning, knowledge, generation, and image-classification benchmarks, with further gains after fine-tuning.

    Source: §1 · p. 3, §8 · p. 21

  • Takeaways & Limitations

    The open models offer a strong foundation for developing downstream healthcare AI applications across medical image and text tasks.

    Source: §8 · p. 21, §9 · p. 22

    The multimodal MedGemma 27B variant has only preliminary evaluation, while broader validation on challenging benchmarks remains ongoing.

    Source: §1 · p. 3, §8 · p. 21

Abstract

from arXiv · show

Artificial intelligence (AI) has significant potential in healthcare applications, but its training and deployment faces challenges due to healthcare's diverse data, complex tasks, and the need to preserve privacy. Foundation models that perform well on medical tasks and require less task-specific tuning data are critical to accelerate the development of healthcare AI applications. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models. For out-of-distribution tasks, MedGemma achieves 2.6-10% improvement on medical multimodal question answering, 15.5-18.1% improvement on chest X-ray finding classification, and 10.8% improvement on agentic evaluations compared to the base models. Fine-tuning MedGemma further improves performance in subdomains, reducing errors in electronic health record information retrieval by 50% and reaching comparable performance to existing specialized state-of-the-art methods for pneumothorax classification and histopathology patch classification. We additionally introduce MedSigLIP, a medically-tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and as an encoder achieves comparable or better performance than specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with potential to significantly accelerate medical research and development of downstream applications. The MedGemma collection, including tutorials and model weights, can be found at https://goo.gle/medgemma.

1. Introduction

MedGemma is an open suite of medically tuned vision-language foundation models built on Gemma 3 to interpret and reason over medical images and text while retaining general capabilities. The collection includes multimodal and text-only models plus the standalone MedSigLIP medical image encoder.

Motivation and design

  • MedGemma addresses generic LMMs’ limited nuanced medical understanding by combining Gemma 3’s architecture with medical image-and-text interpretation and reasoning.The models are designed to retain Gemma 3’s strong general-purpose capabilities while improving robustness on medical data.Source: §1 · p. 2

Released collection

  • The released collection includes MedGemma models, MedSigLIP, tutorials, and download links for the models.Figure 1 provides a high-level overview of the collection, including the MedSigLIP encoder and the MedGemma 4B Multimodal and 27B Text models.Source: Figure 1 · p. 2, §1 · p. 3

Model capabilities

  • MedGemma 4B performs strongly on VQA benchmarks against prior SOTA models such as Med-Gemini, while both 4B and 27B are competitive on challenging text-only medical benchmarks.The text-only evaluations include MedQA, MedMCQA, PubMedQA, MMLU Med, AfriMed-QA, and AgentClinic; fine-tuning further improves chest X-ray reporting, histopathology classification, and EHR information retrieval.Source: §1 · pp. 2–3

Model variants

  • The report focuses on multimodal MedGemma 4B and text-only MedGemma 27B, while evaluation of the released multimodal 27B variant remains ongoing.Unless otherwise noted, “MedGemma 27B” refers to the text-only variant, with preliminary multimodal results reported in Appendix Section F.Source: §1 · pp. 2–3

MedSigLIP

  • MedSigLIP is a standalone 400M-parameter medical image encoder based on SigLIP-400M that enables data-efficient and zero-shot classification and retrieval.MedSigLIP powers MedGemma’s image interpretation and performs comparably to or better than specialized image encoders.Source: §1 · p. 3

2. Methods

MedGemma builds on Gemma 3 using medical datasets, enhanced visual encoding, and staged multimodal training. The methods preserve general-purpose capabilities while adapting the models to diverse 2D medical imaging and text tasks.

2.1.1. Training datasets

  • Training data combined teacher-generated responses and logits from multiple medical QA datasets with approximately 200,000 synthetic medical questions.The source datasets included MedQA, MedMCQA, PubMedQA, MedExpQA, AfriMed-QA, HealthSearchQA, and LiveQA.Source: §2.1.1 · p. 3

2.1. Datasets

  • The dataset strategy extends Med-Gemini and reuses SigLIP and Gemma 3 mixtures while focusing MedGemma’s multimodal training on 2D medical images.3D volumes and genomic datasets were excluded, and additional internal data covered ophthalmology, dermatology, histopathology, and radiology.Source: §2.1.1 · pp. 3–4

2.1.2. Data Preprocessing

  • Images retained prior padding and resizing procedures but used 896×896 inputs, a 262,000-entry SentencePiece tokenizer, and three CT windows mapped to RGB channels.The CT windows highlighted bone and lung, soft tissue, and brain using specified window widths and levels.Source: §2.1.2 · p. 4

2.2.1. Modeling Architecture and Training Infrastructure

  • The architecture follows Gemma 3 with a shared 400M SigLIP vision encoder, 896×896 inputs, arbitrary image-text interleaving, and 128k context.Training used TPUv4, TPUv5e, and TPUv5p with pre-computed visual tokens and multi-pod data and model sharding.Source: §2.2.1 · p. 4

2.2.2. Model Training

  • Training enhanced the vision encoder with over 33M medical image-text pairs, continued multimodal pretraining with medical data weighted at 10%, and applied distillation and reinforcement learning.The 4B multimodal model used all stages, whereas the text-only 27B model used post-training alone; multimodal post-training used RL for better generalization than supervised fine-tuning.Source: §2.2.2 · p. 5

3. MedGemma Evaluations

MedGemma was evaluated across five medical task types and general-purpose benchmarks, using public test data, private medical datasets, and standardized comparison procedures. The evaluation covered text, images, multimodal question answering, report generation, agentic behavior, and potential specialization tradeoffs.

3. MedGemma Evaluations

  • The evaluation covered text question-answering, image classification, visual question answering, chest X-ray report generation, agentic behavior, and general-purpose benchmarks.Table 2 provides a high-level overview of the evaluation tasks and datasets.Source: §3 · p. 5, Table 2 · p. 6

3.1. General evaluation approach

  • Evaluations generally used one inference run per example, with temperature 0.0 for MedGemma on medical benchmarks and default temperatures for non-medical benchmarks and other models.Generalist models were prompted with helpful medical or radiology assistant personas, and DeepSeek R1 was evaluated only on text benchmarks because it is text-only.Source: §3.1 · pp. 5–6
  • Comparisons used publicly documented models meeting criteria for accessible model cards, attribution, licensing, and technical reporting, while VQA comparisons were restricted to zero-shot generative evaluations.Publicly available datasets were required for evaluations involving public APIs, and existing published metrics were used when available.Source: §3.1 · p. 6

3.2. Medical text question-answering

  • Medical text evaluation used official public test splits for MedQA, MedMCQA, PubMedQA, MMLU medical subcategories, AfriMed-QA, and MedXpertQA.AfriMed-QA evaluations used only closed multiple-choice questions, and MedXpertQA was treated as out-of-distribution because it was not used in training.Source: §3.2 · pp. 6–7

3.3. Medical image classification

  • Medical image classification covered public chest X-ray datasets and private dermatology, histopathology, and retinal fundus datasets, using accuracy or macro F1 as appropriate.Chest X-ray tasks targeted specified findings across MIMIC-CXR, CheXpert, and ChestX-ray14, while other datasets represented dermatologic conditions, tissue diagnoses, and diabetic retinopathy severity.Source: §3.3 · p. 7

3.4. Medical visual question-answering

  • Visual question answering used SLAKE and VQA-RAD, measuring average tokenized F1 across open and closed questions and accuracy on yes/no questions.VQA-RAD used contamination-avoiding splits from Yang et al. (2024), while SLAKE used its default train/test split.Source: §3.4 · pp. 7–8

3.5. Chest X-ray report generation

  • Chest X-ray report generation evaluated pretrained MedGemma 4B on MIMIC-CXR using RadGraph F1 against radiologist reports and a five-point human expert comparison.The pretrained model was selected because its reporting style better matched MIMIC-CXR, while human review complemented automated scoring by assessing major and minor issues, omissions, and errors.Source: §3.5 · p. 8

3.6. Medical agentic behavior

  • Agentic evaluation tested MedGemma as a physician agent on 415 simulated AgentClinic cases requiring history taking, exam ordering and interpretation, and diagnosis under incomplete information.The cases comprised AgentClinic-MedQA with 215 cases and AgentClinic-MIMIC-IV with 200 cases.Source: §3.6 · p. 8

4. MedGemma Results

MedGemma consistently outperformed same-sized Gemma 3 baselines across medical text, vision, VQA, report-generation, and agentic evaluations, while retaining general-purpose capabilities. Its strongest advantages appeared in medical vision tasks, where smaller models sometimes surpassed much larger general-purpose systems.

Medical text question-answering

  • MedGemma outperformed same-sized Gemma 3 models across text-only biomedical QA and remained competitive with much larger models on many benchmarks, including MedXpertQA.Evaluated datasets included MedQA, MedMCQA, PubMedQA, MMLU subsets, AfriMed, and out-of-distribution MedXpertQA.Source: §4 · p. 8, Table 3 · p. 9, Table 4 · p. 9

Medical image classification

  • MedGemma achieved substantially superior performance to same-sized Gemma 3 baselines across medical image classification, including radiology, histopathology, dermatology, and ophthalmology, and exceeded larger API models in these tasks.Evaluations used zero-shot generative classification, which may understate performance relative to embedding-based classifiers.Source: §4 · p. 9, Table 7 · p. 14, Table 8 · p. 14

Medical visual question answering

  • MedGemma 4B outperformed same-sized Gemma 3 on SLAKE and VQA-RAD while remaining competitive with much larger models in medical visual question answering.Comparisons exclude discriminative embedding-based approaches that use training data to define possible answers; reported LLaVA-Med and BiomedGPT-B results involved supervised dataset-specific fine-tuning.Source: §4 · pp. 9, 11, Table 9 · p. 15

Chest X-ray report generation

  • 81% of MedGemma-generated chest X-ray reports produced the same or superior clinical decisions as the original reports, while 68% of normal and 49% of abnormal reports were equal or superior to originals.The pretrained models achieved SOTA-level results after training with MIMIC-CXR as the only chest X-ray-related dataset; the human-evaluation result is shown in Figure 4.Source: §4 · p. 11, Figure 4 · p. 13, Table 10 · p. 15

Medical agentic behavior

  • MedGemma 27B outperformed Gemma 3 27B in AgentClinic, exceeded human physicians on AgentClinic-MedQA, and approached much larger models on both evaluated environments.The 4B variants were poorly suited to the agentic task because they had difficulty following AgentClinic system instructions.Source: §4 · p. 11

General-purpose benchmarks and observed themes

  • MedGemma showed only minor decreases from same-sized Gemma 3 models on general-purpose benchmarks, while specialized models delivered a distinct size-to-performance advantage and sometimes surpassed much larger general-purpose models in medical vision.The computational cost difference between MedGemma 4B and the most expensive comparator was 500fold; older benchmarks may also be affected by test-data leakage.Source: §4 · p. 14

5. MedGemma Fine-tuning Demonstration

MedGemma can be adapted to specialized multimodal and longitudinal EHR tasks through supervised fine-tuning or reinforcement learning. Fine-tuning improved task performance, including new state-of-the-art MIMIC-CXR report generation and narrowed MedGemma 27B’s EHRQA gap with top models.

5. MedGemma Fine-tuning Demonstration

  • The demonstration covered three multimodal tasks and one text-only EHR task, using SFT for multimodal adaptation and RL for EHR reasoning.MIMIC-CXR report generation strictly required SFT to capture implicit reporting style, whereas other tasks could use either SFT or RL; trade-offs depended on data size, distribution shift, and task difficulty.Source: §5 · p. 16

5.1. Fine-tuning for multimodal tasks

  • Multimodal fine-tuning used full-parameter SFT with cross-entropy loss, task-specific prompts, validation splits, learning-rate sweeps, and one-epoch training.For report generation, images and indications prompted prediction of findings and impressions; classification tasks were formulated as multiple-choice questions, with input images and prompts excluded from loss computation.Source: §5.1 · pp. 16–17
  • Fine-tuned MedGemma 4B established new state-of-the-art performance, achieving a RadGraph F1 score of 30.3 on the MIMIC-CXR report generation task and approaching SOTA models.Source: §5.1 · p. 17

5.2. Fine-tuning for EHR information retrieval and reasoning

  • EHRQA was newly developed from synthetic FHIR records with programmatic multi-hop reasoning and clinician-designed question types, but it excludes clinical notes.The benchmark contains 81 records split into training, validation, and test sets, with approximately 200 questions per patient and automated evaluation across ten categories.Source: §5.2 · p. 17
  • RL substantially improved MedGemma 27B’s EHRQA accuracy, closing its gap with top-performing models, especially for reasoning across inter-dependent records.The EHRQA benchmark uses longitudinal synthetic FHIR records and evaluates ten question categories; its dataset lacks clinical notes.Source: §5.2 · pp. 17–18, Table 14 · p. 18

6. MedSigLIP Evaluations

MedSigLIP was evaluated through zero-shot and linear-probe classification across chest X-ray, dermatology, ophthalmology, and histopathology datasets. Zero-shot evaluation used prompt-based image–text similarity and AUC, while linear probing trained logistic regression on MedSigLIP image embeddings.

Evaluation datasets

  • MedSigLIP evaluations covered four modalities—chest X-ray, dermatology, ophthalmology, and histopathology—using established datasets and task-specific class sets.Chest X-ray evaluations used CheXpert, CXR14, and ELIXR frameworks; dermatology used a 79-condition, 1,612-patient US-Derm MCQA subset; ophthalmology used EyePACS; histopathology used patches from whole-slide images.Source: §6 · p. 18

Zero-shot evaluation

  • Zero-shot classification averaged multiple class-prompt embeddings, ranked image–class cosine similarities with softmax, and reported 1-vs-all AUC for multiclass conditions.Dermatology prompts were condition names, while ophthalmology, chest X-ray, and histopathology prompts were specified in appendix tables.Source: §6 · p. 18

Linear-probe evaluation

  • Linear-probe classification extracted MedSigLIP image embeddings without its text encoder and trained SAGA logistic regression with validation-selected hyperparameters before test evaluation.This data-efficient setting measured performance after additional training on target-condition labels.Source: §6 · p. 18

7. MedSigLIP Results

MedSigLIP provides strong single-model medical image classification across chest X-ray, dermatology, ophthalmology, and histopathology. It often matches or exceeds domain-specific foundation models, while task-specific training further improves performance.

Chest X-ray

  • 2.0% higher zero-shot CXR performance than CXR Foundation, with fractures improving over ELIXR by 7.1% despite MedSigLIP’s lower image resolution.MedSigLIP also performed strongly for linear probing when training sets contained at least 512 examples across seven findings.Source: §7 · p. 19, Figure 5 · p. 19, Table 16 · p. 20

Dermatology

  • MedSigLIP’s zero-shot and linear-probe performance outperformed Derm Foundation linear probes for distinguishing 79 skin conditions.Derm Foundation and Path Foundation could not be evaluated zero-shot because they are image-only comparator models.Source: §7 · p. 19, Table 15 · p. 20

Ophthalmology

  • 9.8% higher linear-probe than zero-shot performance on five-class diabetic-retinopathy classification, with no HAI-DEF ophthalmology model available for comparison.The five classes were none, mild, moderate, severe, and proliferative diabetic retinopathy.Source: §7 · p. 19, Table 15 · p. 20

Histopathology

  • Histopathology linear probing modestly improved over MedSigLIP zero-shot classification and approached Path Foundation, whose linear-probe performance was 1.9% better on average.These results compare AUCs for histopathology findings across MedSigLIP and HAI-DEF image models.Source: §7 · p. 19, Table 15 · p. 20

Overall findings

  • MedSigLIP offers strong zero-shot baseline performance across medical domains, while task-specific images and logistic-regression training can yield stronger results.It is a single model evaluated against dedicated domain-specific HAI-DEF image models, including Derm Foundation and Path Foundation.Source: §7 · pp. 19–20

8. Discussion

MedGemma and MedSigLIP are open medical models built on Gemma 3 that show strong benchmark performance, improve with fine-tuning, and support adaptable healthcare applications. Their real-world utility remains to be validated beyond current benchmarks through further evaluation and application development.

Contributions

  • MedGemma and MedSigLIP achieved strong performance across clinical reasoning, biomedical knowledge, report generation, and medical image classification, with further gains after fine-tuning.The models were optimized for medical domains and are intended as starting points for developing healthcare AI applications.Source: §8 · p. 21

Advantages over general models

  • Domain-specific pre-training and post-training improved MedGemma over base Gemma 3 across evaluated benchmarks and brought performance on par with much larger models.These advantages distinguish MedGemma from general models for developers seeking medical specialization without relying on larger systems.Source: §8 · p. 21
  • MedGemma is suited to frozen, locally deployable, cost-sensitive, and adaptable medical applications, while larger models remain preferable for optimal broad performance or complementary agentic use.The comparison with general API-based models depends on deployment constraints, medical capabilities, and control over adaptation.Source: §8 · p. 21

Downstream applications

  • MedGemma can support medical image retrieval, multimodal data integration, clinical-text extraction, workflow assistance, patient communication, and customized agentic healthcare systems.Potential applications include cohort creation, educational tools, clinical-trial matching, pharmacovigilance, healthcare-quality analysis, and radiology support.Source: §8 · p. 21

Limitations and future work

  • Automated benchmarks are only an initial validation step, and some may be near saturation; further work should test challenging real-world benchmarks and application or agentic integration (Alaa et al., 2025).These evaluations are intended to guide optimal use cases and future model versions with extended capabilities.Source: §8 · p. 21

Open release

  • The open release of MedGemma and MedSigLIP enables community evaluation, improvement, and adaptation, providing predictability and flexibility for healthcare developers.The authors intend this openness to accelerate AI application development across diverse healthcare use cases.Source: §8 · p. 21

9. Conclusion

MedGemma demonstrates robust capabilities across diverse medical vision-language and text-only tasks, while MedSigLIP provides robust multi-domain capabilities as a medical foundation model. Their breadth and efficiency support potential medical applications, contingent on careful validation of safety, performance, and reliability.

Conclusion

  • MedGemma demonstrates robust capabilities across a variety of vision-language and text-only medical tasks.Source: §9 · p. 22
  • MedSigLIP demonstrates robust multi-domain capabilities and can serve as a strong medical foundation model.Source: §9 · p. 22
  • The models’ breadth and efficiency offer possibilities for diverse use cases, but downstream applications require thoughtful validation of safety, performance, and reliability.Source: §9 · p. 22
  • Providing MedGemma and MedSigLIP to developers under a permissive license aims to enable useful and innovative medical applications.Source: §9 · p. 22

11. Contributions and Acknowledgments

The Contributions subsection acknowledges the paper’s contributors and identifies special authorship roles through co-first and co-last author notation.

Contributions

  • The section lists the full group of contributors across multiple entries, including Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, and the remaining named authors.The contributor list spans all seven contribution passages.Source: Contributions · p. 23
  • ∗ marks co-first authors, while † marks co-last authors.Rory Pilgrim† appears in the contributor list, and the notation is defined separately.Source: Contributions · p. 23

Use of AI in Manuscript Preparation · Appendix · A. Manual Evaluation of Radiology reports

The manuscript introduction was drafted manually, refined with Gemini 2.5 Pro, and manually checked for accuracy. Appendix A provides a rubric for radiologists evaluating AI-generated chest X-ray reports against original reports.

Use of AI in Manuscript Preparation

  • The introduction was drafted manually, refined using Gemini 2.5 Pro, and checked manually for content accuracy.The authors state that they take full responsibility for the content.Source: Use of AI in Manuscript Preparation · p. 24

A. Manual Evaluation of Radiology reports

B. Evaluation Prompts · C. Comparison of CXR data-efficient learning · D. Additional details about EHRQA

The appendix specifies evaluation prompts and system-temperature settings, compares MedSigLIP with ELIXR for data-efficient chest X-ray learning, and details EHRQA comparisons, categories, and evaluation prompts.

B. Evaluation Prompts

  • MedSigLIP zero-shot evaluations use task-specific prompts covering class definitions and ophthalmology, CXR, histopathology, and text-based medical evaluations.These prompts are listed in Tables A2–A7, including image-classification and VQA system-message conventions.Source: B. Evaluation Prompts · p. 33, Table A2 · p. 33, Table A3 · p. 34
  • MedGemma evaluation uses temperature 0.0, while other models use their default temperatures and receive radiology- or medical-assistant system prompts.Per-dataset prompts and model-specific system-message prefixes are provided in Appendix Tables A6 and A7.Source: B. Evaluation Prompts · p. 33, Table A7 · p. 49
  • Histopathology prompts span graded breast and cervical lesions, prostate Gleason patterns, TCGA cancer types, and normal tissue categories.The prompt sets encode distinctions such as nucleopleomorphism, tubule formation, dysplasia, glandular architecture, and tissue origin.Source: B. Evaluation Prompts · pp. 37–40

C. Comparison of CXR data-efficient learning

D. Additional details about EHRQA

E. Additional medical reasoning examples · F. MedGemma 27B multimodal model

The appendix provides additional MedGemma 27B clinical-reasoning examples with specialist commentary and introduces a 27B multimodal variant with expanded training and preliminary evaluation. The examples show structured reasoning while also identifying dosing, diagnostic, and clinical-context limitations.

E. Additional medical reasoning examples

  • In a cardiology example, MedGemma correctly identifies sacubitril/valsartan as indicated for symptomatic HFrEF but appropriately flags ACE-inhibitor washout, monitoring, and contraindication considerations.The reasoning uses EF, symptoms, blood pressure, renal function, potassium, and concomitant medications to justify the recommendation.Source: Table A10 · p. 55
  • Cardiology commentary finds the reasoning strong overall but notes an incorrect lowest dose, reliance on creatinine rather than eGFR, insufficient carvedilol consideration, and omission of empagliflozin.The commentary also notes that carvedilol titration could reduce blood-pressure room for Entresto and requires careful concurrent titration.Source: E. Additional medical reasoning examples · p. 56
  • Table A11 presents an example of MedGemma 27B open clinical reasoning in a cardiology setting with succinct output and commentary by a US board-certified practicing cardiologist.Source: Table A11 · p. 57
  • In an infectious-disease case, the model prioritizes leptospirosis from the travel history and multisystem presentation, while considering dengue, hantavirus, rickettsial infection, and other alternatives.Its differential reasoning connects jaundice, acute kidney injury, thrombocytopenia, pulmonary disease, and possible myocarditis to the leading diagnosis.Source: Table A12 · p. 57, E. Additional medical reasoning examples · p. 58
  • The infectious-disease reasoning recommends serologic and PCR testing for leptospirosis, dengue, and hantavirus, including urine PCR for leptospirosis given the symptom duration.It also recommends monitoring organ function and considering broader testing if the initial workup is negative or the clinical picture changes.Source: E. Additional medical reasoning examples · p. 58
  • Specialist commentary praises the infectious-disease answer’s structure and focus but challenges its Hawaii epidemiology and overly narrow testing, emphasizing broader evaluation and possible empiric treatment during laboratory delays.The commentary notes that several tests may require specialized laboratories and courier transport, delaying results.Source: E. Additional medical reasoning examples · p. 59

F. MedGemma 27B multimodal model

  • The 27B multimodal variant retains the 4B model’s capabilities while improving language, EHR understanding, and chest-X-ray anatomy localization; formal evaluation remains ongoing.It adds EHRQA and Chest ImaGenome training data, with preliminary results reported in Table A14.Source: Table A14 · p. 60, F. MedGemma 27B multimodal model · p. 60

Related papers

Loading 2507.05201v4…