{
  "id": "fd5f91a5-3210-460e-8bc4-ff8d0aa056b8",
  "name": "Open-MAI-Dx-Orchestrator",
  "title": "Agent",
  "description": "An open source implementation of the paper: \"Sequential Diagnosis with Language Models\" From Microsoft Built with Swarms Framework\n\n# MAI Diagnostic Orchestrator (MAI-DxO)\n\n> **An open-source implementation of Microsoft Research's \"Sequential Diagnosis with Language Models\" paper, built with the Swarms AI framework.**\n\nMAI-DxO (MAI Diagnostic Orchestrator) is a sophisticated AI-powered diagnostic system that simulates a virtual panel of physician-agents to perform iterative medical diagnosis with cost-effectiveness optimization. This implementation faithfully reproduces the methodology described in the Microsoft Research paper while providing additional features and flexibility.\n\n[![Paper](https://img.shields.io/badge/Paper-arXiv:2506.22405-red.svg)](https://arxiv.org/abs/2506.22405)\n[![License](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)\n[![Python](https://img.shields.io/badge/Python-3.8+-green.svg)](https://python.org)\n\n## ✨ Key Features\n\n- **8 AI Physician Agents**: Specialized roles for comprehensive diagnosis.\n- **5 Operational Modes**: Instant, question-only, budgeted, no-budget, and ensemble modes.\n- **Cost Tracking**: Real-time budget monitoring with costs for 25+ medical tests.\n- **Clinical Evaluation**: 5-point accuracy scoring with detailed feedback.\n- **Model Agnostic**: Works with GPT, Gemini, Claude, and other leading LLMs.\n- **Token-Optimized Prompts**: Ultra-compact role prompts reduce token usage and latency without sacrificing reasoning quality.\n\n## 🚀 Quick Start\n\n### 1. Installation\n\nInstall the package directly via pip:\n\n```bash\npip install mai-dx\n```\n\nOr, for development, clone the repository and install the requirements:\n\n```bash\ngit clone https://github.com/The-Swarm-Corporation/Open-MAI-Dx-Orchestrator.git\ncd Open-MAI-Dx-Orchestrator\npip install -r requirements.txt\n```\n\n### 2. Environment Setup\n\nCreate a `.env` file in your project root and add your API keys:\n\n```txt\nOPENAI_API_KEY=\"Your OpenAI API key\"\nGEMINI_API_KEY=\"Your Gemini API key\"\nANTHROPIC_API_KEY=\"Your Anthropic API key\"\n```\n\n### 3. Basic Usage\n\n```python\nfrom mai_dx import MaiDxOrchestrator\n\n# Create the orchestrator (defaults to a capable model)\norchestrator = MaiDxOrchestrator()\n\n# Run a diagnosis\nresult = orchestrator.run(\n    initial_case_info=\"29-year-old woman with sore throat and peritonsillar swelling...\",\n    full_case_details=\"Patient: 29-year-old female. History: Onset of sore throat...\",\n    ground_truth_diagnosis=\"Embryonal rhabdomyosarcoma of the pharynx\"\n)\n\n# Print the results\nprint(f\"Final Diagnosis: {result.final_diagnosis}\")\nprint(f\"Accuracy: {result.accuracy_score}/5.0\")\nprint(f\"Total Cost: ${result.total_cost:,.2f}\")\n```\n\n## ⚙️ Advanced Usage & Configuration\n\nCustomize the orchestrator's model, budget, and operational mode.\n\n```python\nfrom mai_dx import MaiDxOrchestrator\n\n# Configure with a specific model and budget\norchestrator = MaiDxOrchestrator(\n    model_name=\"gemini/gemini-2.5-flash\",  # or \"gpt-4\", \"claude-3-5-sonnet\"\n    max_iterations=10,\n    initial_budget=3000,\n    mode=\"budgeted\"  # Other modes: \"instant\", \"question_only\", \"no_budget\"\n)\n\n# Run the diagnosis\n# ...\n```\n\n## 🏥 How It Works: The Virtual Physician Panel\n\nMAI-DxO employs a multi-agent system where each agent has a specific role:\n\n- **🧠 Dr. Hypothesis**: Maintains the differential diagnosis.\n- **🔬 Dr. Test-Chooser**: Selects the most cost-effective diagnostic tests.\n- **🤔 Dr. Challenger**: Prevents cognitive biases and diagnostic errors.\n- **💰 Dr. Stewardship**: Ensures cost-effective care.\n- **✅ Dr. Checklist**: Performs quality control checks.\n- **🤝 Consensus Coordinator**: Synthesizes panel decisions.\n- **🔑 Gatekeeper**: Acts as the clinical information oracle.\n- **⚖️ Judge**: Evaluates the final diagnostic accuracy.\n\n\n## Documentation\n\nLearn more about this repository [with the docs](DOCS.md)\n\n## 🤝 Contributing\n\nWe welcome contributions! Please feel free to open an issue or submit a pull request.\n\n## 📄 License\n\nThis project is licensed under the MIT License. See the [LICENSE](LICENSE) file for details.\n\n## 📚 Citation\n\nIf you use this work in your research, please cite both the original paper and this software implementation.\n\n```bibtex\n@misc{nori2025sequentialdiagnosislanguagemodels,\n    title={Sequential Diagnosis with Language Models}, \n    author={Harsha Nori and Mayank Daswani and Christopher Kelly and Scott Lundberg and Marco Tulio Ribeiro and Marc Wilson and Xiaoxuan Liu and Viknesh Sounderajah and Jonathan Carlson and Matthew P Lungren and Bay Gross and Peter Hames and Mustafa Suleyman and Dominic King and Eric Horvitz},\n    year={2025},\n    eprint={2506.22405},\n    archivePrefix={arXiv},\n    primaryClass={cs.CL},\n    url={https://arxiv.org/abs/2506.22405}, \n}\n\n@software{mai_dx_orchestrator,\n    title={Open-MAI-Dx-Orchestrator: An Open Source Implementation of Sequential Diagnosis with Language Models},\n    author={The-Swarm-Corporation},\n    year={2025},\n    url={https://github.com/The-Swarm-Corporation/Open-MAI-Dx-Orchestrator.git}\n}\n```\n\n## 🔗 Related Work\n\n- [Original Paper](https://arxiv.org/abs/2506.22405) - Sequential Diagnosis with Language Models\n- [Swarms Framework](https://github.com/kyegomez/swarms) - Multi-agent AI orchestration\n- [Microsoft Research](https://www.microsoft.com/en-us/research/) - Original research institution\n\n## 📞 Support\n\n- **Issues**: [GitHub Issues](https://github.com/The-Swarm-Corporation/Open-MAI-Dx-Orchestrator/issues)\n- **Discussions**: [GitHub Discussions](https://github.com/The-Swarm-Corporation/Open-MAI-Dx-Orchestrator/discussions)\n- **Documentation**: [Full Documentation](https://docs.swarms.world)\n\n---\n\n<p align=\"center\">\n  <strong>Built with Swarms for advancing AI-powered medical diagnosis</strong>\n</p>\n\nSource: https://github.com/The-Swarm-Corporation/Open-MAI-Dx-Orchestrator",
  "tags": [
    "ai",
    "microsoft",
    "ml",
    "multi-agent",
    "paper-implementation",
    "swarms",
    "swarms-agents",
    "swarms-ecosystem",
    "ehr",
    "epic",
    "medical-agents",
    "healthcare-agennts",
    "xray-agents"
  ],
  "requirements": [
    {
      "package": "swarms",
      "installation": "pip3 install swarms"
    }
  ],
  "usecases": [
    {
      "title": "",
      "description": ""
    }
  ],
  "userId": "12fa3db5-0635-41b6-aa8c-71ece35a18db",
  "createdAt": "2025-07-11T17:03:18.568173+00:00",
  "links": [],
  "code": "from mai_dx import MaiDxOrchestrator\nfrom loguru import logger\n\nif __name__ == \"__main__\":\n    # Example case inspired by the paper's Figure 1\n    initial_info = (\n        \"A 29-year-old woman was admitted to the hospital because of sore throat and peritonsillar swelling \"\n        \"and bleeding. Symptoms did not abate with antimicrobial therapy.\"\n    )\n\n    full_case = \"\"\"\n    Patient: 29-year-old female.\n    History: Onset of sore throat 7 weeks prior to admission. Worsening right-sided pain and swelling.\n    No fevers, headaches, or gastrointestinal symptoms. Past medical history is unremarkable. No history of smoking or significant alcohol use.\n    Physical Exam: Right peritonsillar mass, displacing the uvula. No other significant findings.\n    Initial Labs: FBC, clotting studies normal.\n    MRI Neck: Showed a large, enhancing mass in the right peritonsillar space.\n    Biopsy (H&E): Infiltrative round-cell neoplasm with high nuclear-to-cytoplasmic ratio and frequent mitotic figures.\n    Biopsy (Immunohistochemistry for Carcinoma): CD31, D2-40, CD34, ERG, GLUT-1, pan-cytokeratin, CD45, CD20, CD3 all negative. Ki-67: 60% nuclear positivity.\n    Biopsy (Immunohistochemistry for Rhabdomyosarcoma): Desmin and MyoD1 diffusely positive. Myogenin multifocally positive.\n    Biopsy (FISH): No FOXO1 (13q14) rearrangements detected.\n    Final Diagnosis from Pathology: Embryonal rhabdomyosarcoma of the pharynx.\n    \"\"\"\n\n    ground_truth = \"Embryonal rhabdomyosarcoma of the pharynx\"\n\n    # --- Demonstrate Different MAI-DxO Variants ---\n    try:\n        print(\"\\n\" + \"=\" * 80)\n        print(\n            \"    MAI DIAGNOSTIC ORCHESTRATOR (MAI-DxO) - SEQUENTIAL DIAGNOSIS BENCHMARK\"\n        )\n        print(\n            \"                    Implementation based on the NEJM Research Paper\"\n        )\n        print(\"=\" * 80)\n\n        # Test different variants as described in the paper\n        variants_to_test = [\n            (\n                \"no_budget\",\n                \"Standard MAI-DxO with no budget constraints\",\n            ),\n            (\"budgeted\", \"Budget-constrained MAI-DxO ($3000 limit)\"),\n            (\n                \"question_only\",\n                \"Question-only variant (no diagnostic tests)\",\n            ),\n        ]\n\n        results = {}\n\n        for variant_name, description in variants_to_test:\n            print(f\"\\n{'='*60}\")\n            print(f\"Testing Variant: {variant_name.upper()}\")\n            print(f\"Description: {description}\")\n            print(\"=\" * 60)\n\n            # Create the variant\n            if variant_name == \"budgeted\":\n                orchestrator = MaiDxOrchestrator.create_variant(\n                    variant_name,\n                    budget=3000,\n                    model_name=\"gemini/gemini-2.5-flash\",\n                    max_iterations=5,\n                )\n            else:\n                orchestrator = MaiDxOrchestrator.create_variant(\n                    variant_name,\n                    model_name=\"gemini/gemini-2.5-flash\",\n                    max_iterations=5,\n                )\n\n            # Run the diagnostic process\n            result = orchestrator.run(\n                initial_case_info=initial_info,\n                full_case_details=full_case,\n                ground_truth_diagnosis=ground_truth,\n            )\n\n            results[variant_name] = result\n\n            # Display results\n            print(f\"\\n🚀 Final Diagnosis: {result.final_diagnosis}\")\n            print(f\"🎯 Ground Truth: {result.ground_truth}\")\n            print(f\"⭐ Accuracy Score: {result.accuracy_score}/5.0\")\n            print(f\"   Reasoning: {result.accuracy_reasoning}\")\n            print(f\"💰 Total Cost: ${result.total_cost:,}\")\n            print(f\"🔄 Iterations: {result.iterations}\")\n            print(f\"⏱️  Mode: {orchestrator.mode}\")\n\n        # Demonstrate ensemble approach\n        print(f\"\\n{'='*60}\")\n        print(\"Testing Variant: ENSEMBLE\")\n        print(\n            \"Description: Multiple independent runs with consensus aggregation\"\n        )\n        print(\"=\" * 60)\n\n        ensemble_orchestrator = MaiDxOrchestrator.create_variant(\n            \"ensemble\",\n            model_name=\"gemini/gemini-2.5-flash\",\n            max_iterations=3,  # Shorter iterations for ensemble\n        )\n\n        ensemble_result = ensemble_orchestrator.run_ensemble(\n            initial_case_info=initial_info,\n            full_case_details=full_case,\n            ground_truth_diagnosis=ground_truth,\n            num_runs=2,  # Reduced for demo\n        )\n\n        results[\"ensemble\"] = ensemble_result\n\n        print(\n            f\"\\n🚀 Ensemble Diagnosis: {ensemble_result.final_diagnosis}\"\n        )\n        print(f\"🎯 Ground Truth: {ensemble_result.ground_truth}\")\n        print(\n            f\"⭐ Ensemble Score: {ensemble_result.accuracy_score}/5.0\"\n        )\n        print(\n            f\"💰 Total Ensemble Cost: ${ensemble_result.total_cost:,}\"\n        )\n\n        # --- Summary Comparison ---\n        print(f\"\\n{'='*80}\")\n        print(\"                           RESULTS SUMMARY\")\n        print(\"=\" * 80)\n        print(\n            f\"{'Variant':<15} {'Diagnosis Match':<15} {'Score':<8} {'Cost':<12} {'Iterations':<12}\"\n        )\n        print(\"-\" * 80)\n\n        for variant_name, result in results.items():\n            match_status = (\n                \"✓ Match\"\n                if result.accuracy_score >= 4.0\n                else \"✗ No Match\"\n            )\n            print(\n                f\"{variant_name:<15} {match_status:<15} {result.accuracy_score:<8.1f} ${result.total_cost:<11,} {result.iterations:<12}\"\n            )\n\n        print(f\"\\n{'='*80}\")\n        print(\n            \"Implementation successfully demonstrates the MAI-DxO framework\"\n        )\n        print(\n            \"as described in 'Sequential Diagnosis with Language Models' paper\"\n        )\n        print(\"=\" * 80)\n\n    except Exception as e:\n        logger.exception(\n            f\"An error occurred during the diagnostic session: {e}\"\n        )\n        print(f\"\\n❌ Error occurred: {e}\")\n        print(\"Please check your model configuration and API keys.\")\n"
}