toolevalxm/MedDiagnoseAI-BestCheckpoint

MedDiagnoseAI

---

1. Introduction

MedDiagnoseAI represents a breakthrough in clinical decision support systems. The latest version has been trained on over 50 million de-identified patient records, incorporating multi-modal data including clinical notes, lab results, imaging reports, and genomic markers. The model demonstrates exceptional performance across various clinical benchmarks, approaching the diagnostic accuracy of board-certified physicians in many domains.

Compared to the previous version, MedDiagnoseAI v2.0 shows significant improvements in differential diagnosis tasks. For instance, in the MIMIC-IV diagnostic challenge, the model's F1-score has increased from 0.72 in the previous version to 0.89 in the current version. This advancement stems from enhanced clinical context understanding: the new version processes an average of 8K tokens per patient case, compared to 3K tokens in the previous version.

Beyond improved diagnostic capabilities, this version offers reduced false positive rates and enhanced support for multi-specialty consultations.

2. Evaluation Results

Comprehensive Clinical Benchmark Results

	Benchmark	ModelA	ModelB	ModelA-v2	MedDiagnoseAI
Core Diagnostic Tasks	Diagnosis Accuracy	0.620	0.645	0.658	0.837
	Clinical Reasoning	0.701	0.718	0.735	0.779
	Medical Knowledge	0.752	0.769	0.781	0.755
Imaging & Analysis	Radiology Interpretation	0.589	0.612	0.628	0.733
	Patient Q&A	0.634	0.651	0.667	0.633
	Disease Classification	0.812	0.829	0.841	0.853
	Symptom Severity	0.723	0.738	0.752	0.717
Treatment Tasks	Treatment Planning	0.567	0.589	0.604	0.702
	Clinical Documentation	0.645	0.662	0.678	0.645
	Patient Interaction	0.698	0.715	0.729	0.685
	Medical Summarization	0.756	0.772	0.785	0.780
Specialized Capabilities	Medical Terminology	0.834	0.849	0.861	0.840
	Literature Retrieval	0.612	0.631	0.648	0.612
	Protocol Adherence	0.689	0.708	0.723	0.728
	Drug Safety	0.778	0.795	0.812	0.823

Overall Performance Summary

MedDiagnoseAI demonstrates strong performance across all evaluated clinical benchmark categories, with particularly notable results in diagnostic reasoning and drug safety evaluation tasks.

3. Clinical Dashboard & API Platform

We offer a HIPAA-compliant clinical dashboard and API for healthcare institutions to integrate MedDiagnoseAI. Please check our official website for more details.

4. How to Deploy Locally

Please refer to our deployment guide for information about running MedDiagnoseAI in your clinical environment.

Compared to previous versions, the deployment recommendations for MedDiagnoseAI have the following changes:

1. HIPAA-compliant audit logging is now supported by default.
2. Multi-institution federated inference is available without additional configuration.

The model architecture of MedDiagnoseAI-Light is optimized for edge deployment, while sharing the same clinical vocabulary as the main MedDiagnoseAI.

System Prompt

We recommend using the following clinical system prompt:

You are MedDiagnoseAI, a clinical decision support assistant.
Current institution: {institution_name}
Date: {current_date}
IMPORTANT: All outputs require physician review before clinical action.

For example,

You are MedDiagnoseAI, a clinical decision support assistant.
Current institution: Johns Hopkins Hospital
Date: May 28, 2025, Monday.
IMPORTANT: All outputs require physician review before clinical action.

Temperature

We recommend setting the temperature parameter $T_{model}$ to 0.3 for clinical applications to ensure consistent and reliable outputs.

Prompts for Patient Data Input

For patient record analysis, please follow the template to create prompts, where {patient_id}, {clinical_data} and {query} are arguments.

patient_template = \
"""[Patient ID]: {patient_id}
[Clinical Data Begin]
{clinical_data}
[Clinical Data End]
Clinical Query: {query}"""

For literature-enhanced clinical reasoning, we recommend the following prompt template where {pubmed_results}, {cur_date}, and {clinical_question} are arguments.

literature_search_template = \
'''# The following are relevant medical literature findings:
{pubmed_results}
In the literature I provide, each article is formatted as [Article X begin]...[Article X end]. Please cite evidence when making clinical recommendations using [citation:X] format. Multiple citations should be listed as [citation:3][citation:5].

When providing clinical guidance:
- Today is {cur_date}.
- Filter literature by relevance to the specific clinical scenario.
- Prioritize recent systematic reviews and RCTs over case reports.
- For treatment recommendations, include level of evidence.
- Always note when evidence is limited or conflicting.
- Include relevant contraindications and drug interactions.
- Synthesize findings from multiple sources when applicable.
# Clinical question:
{clinical_question}'''

5. License

This code repository is licensed under the Apache 2.0 License. The use of MedDiagnoseAI models requires compliance with healthcare data regulations in your jurisdiction. The model is NOT approved for autonomous clinical decision-making.

6. Contact

For research collaborations or institutional licensing, please contact us at medical@meddiagnoseai.health or submit an inquiry through our compliance portal.