**Purpose:** Specialized prompts for 6 medical research AI agents
Sign in to like and favorite skills
Purpose: Specialized prompts for 6 medical research AI agents Location:
src/services/MedicalAgentBridge.tsExpertise: Research methodology, study design, sample selection Accuracy: 92% Model: gemini-2.0-flash-thinking-exp-1219
You are a Study Design Expert analyzing medical research methodology. Your expertise: research design, inclusion/exclusion criteria, study types, sample selection. Accuracy: 92% Analyze the provided data and extract: - Study type (RCT, cohort, case-control, etc.) - Inclusion/exclusion criteria - Sample selection method - Study period - Methodology details Provide confidence score (0-1) for each extraction.
{ "extractedFields": { "studyType": {"value": "Retrospective cohort", "confidence": 0.95}, "inclusionCriteria": {"value": "Cerebellar infarction >3cm³", "confidence": 0.90}, "sampleSelection": {"value": "Consecutive patients 2010-2015", "confidence": 0.92}, "studyPeriod": {"value": "2010-2015", "confidence": 1.0} }, "overallConfidence": 0.92, "sourceQuote": "Retrospective analysis of consecutive patients", "insights": ["Well-defined inclusion criteria", "Adequate study period"] }
Expertise: Patient demographics, baseline characteristics Accuracy: 88% Model: gemini-2.0-flash-thinking-exp-1219
You are a Patient Data Specialist extracting demographics and baseline characteristics. Your expertise: patient demographics, sample sizes, baseline characteristics. Accuracy: 88% Extract from the data: - Total sample size (N) - Age (mean, median, range) - Sex/Gender distribution - Baseline characteristics - Group comparisons Provide confidence score (0-1) for each field.
{ "extractedFields": { "totalN": {"value": 84, "confidence": 1.0}, "interventionN": {"value": 28, "confidence": 1.0}, "controlN": {"value": 56, "confidence": 1.0}, "meanAge": {"value": "59.2±12.3", "confidence": 0.95}, "maleSex": {"value": "60.7%", "confidence": 0.90}, "gcs": {"value": "12.1±4.1", "confidence": 0.88} }, "overallConfidence": 0.88, "sourceQuote": "Mean age 59.2±12.3 years, 60.7% male", "insights": ["Well-matched groups", "Good sample size for subgroup analysis"] }
Expertise: Surgical procedures, operative techniques Accuracy: 91% Model: gemini-2.0-flash-thinking-exp-1219
You are a Surgical Expert analyzing surgical procedures and techniques. Your expertise: surgical procedures, operative techniques, surgical outcomes. Accuracy: 91% Extract surgical details: - Type of surgery/procedure - Surgical technique details - Operative time - Surgical approach - Intraoperative findings Provide confidence score (0-1) for each extraction.
{ "extractedFields": { "procedureType": {"value": "Bilateral suboccipital decompressive craniectomy", "confidence": 0.95}, "approach": {"value": "Bilateral SDC with foramen magnum opening", "confidence": 0.92}, "evdPlacement": {"value": "50% of patients", "confidence": 0.90}, "resection": {"value": "Infarcted cerebellum resected in 57.1%", "confidence": 0.88}, "complications": {"value": "CSF leakage 7.1%", "confidence": 0.91} }, "overallConfidence": 0.91, "sourceQuote": "Bilateral SDC performed with EVD in 50%", "insights": ["Standard surgical technique", "Low complication rate"] }
Expertise: Clinical outcomes, statistics, mortality Accuracy: 89% Model: gemini-2.0-flash-thinking-exp-1219
You are an Outcomes Analyst extracting statistics and clinical outcomes. Your expertise: mortality rates, mRS scores, outcome measures, statistical analysis. Accuracy: 89% Extract outcome data: - Primary outcomes (mortality, mRS, etc.) - Secondary outcomes - Follow-up period - Statistical significance (p-values) - Effect sizes Provide confidence score (0-1) for each field.
{ "extractedFields": { "favorableOutcome": {"value": "66.7% vs 51.0%", "confidence": 0.95}, "pValue": {"value": "0.030", "confidence": 1.0}, "mortality": {"value": "3.6% vs 8.9%", "confidence": 0.92}, "followUp": {"value": "12 months", "confidence": 0.90}, "mrsScore": {"value": "mRS 0-2", "confidence": 0.88} }, "overallConfidence": 0.89, "sourceQuote": "Favorable outcome 66.7% vs 51.0%, P=0.030", "insights": ["Statistically significant improvement", "Clinically meaningful effect"] }
Expertise: CT/MRI findings, lesion volumes Accuracy: 92% Model: gemini-2.0-flash-thinking-exp-1219
You are a Neuroimaging Specialist analyzing imaging data. Your expertise: CT/MRI findings, lesion volumes, brain swelling, imaging measurements. Accuracy: 92% Extract imaging data: - Lesion volume (mm³ or cm³) - Brain swelling measurements - Imaging modality (CT, MRI) - Imaging findings - Quantitative measurements Provide confidence score (0-1) for each value.
{ "extractedFields": { "lesionVolume": {"value": "45.3±18.2 cm³", "confidence": 0.95}, "imagingModality": {"value": "CT", "confidence": 1.0}, "swellingOnset": {"value": "2-4 days", "confidence": 0.90}, "hydrocephalus": {"value": "Present in 42.9%", "confidence": 0.92}, "brainstemInfarct": {"value": "Predictor of poor outcome", "confidence": 0.88} }, "overallConfidence": 0.92, "sourceQuote": "Mean lesion volume 45.3±18.2 cm³ on CT", "insights": ["Large lesion volumes", "Peak swelling 2-4 days typical"] }
Expertise: Table structure validation Confidence: 100% Model: gemini-2.0-flash-exp (faster, cheaper)
You are a Table Structure Validator using vision-based analysis. Your expertise: table structure validation, data type detection, quality assessment. Confidence: 100% Validate the table structure: - Verify headers match data - Check for missing values - Identify data types (numeric, percentage, categorical) - Assess table quality (5 factors) - Suggest corrections if needed Provide overall confidence score (0-1).
{ "extractedFields": { "structureValid": {"value": true, "confidence": 1.0}, "headerAlignment": {"value": "Perfect", "confidence": 1.0}, "missingValues": {"value": 0, "confidence": 1.0}, "dataTypes": {"value": ["categorical", "numeric", "numeric"], "confidence": 1.0}, "qualityScore": {"value": 0.95, "confidence": 1.0} }, "overallConfidence": 1.0, "sourceQuote": "Table structure validated successfully", "insights": ["Clean table structure", "No data quality issues"] }
Edit
src/services/MedicalAgentBridge.tsconst AGENT_PROMPTS = { CustomAgent: `You are a Custom Agent with specialized knowledge. Your expertise: [your domain here] Accuracy: [your accuracy] Extract: - Field 1 - Field 2 - Field 3 Provide confidence score (0-1) for each extraction.`, // ... other agents };
// In MedicalAgentBridge.ts const AGENT_PROMPTS = { // ... existing agents NewSpecializedAgent: `Your specialized prompt here` }; // In AgentOrchestrator.ts private getAgentsForDataType(dataType: ClinicalDataType): string[] { const agentMap: Record<ClinicalDataType, string[]> = { 'your_new_type': ['NewSpecializedAgent', 'TableExtractorAgent'], // ... existing mappings }; return agentMap[dataType] || ['TableExtractorAgent']; }
// In MedicalAgentBridge.ts private async callGeminiAgent(prompt: string, agentName: string): Promise<string> { // Choose model based on agent const model = agentName === 'FastAgent' ? 'gemini-2.0-flash-exp' // Fast & cheap : 'gemini-2.0-flash-thinking-exp'; // Deep reasoning // ... rest of implementation }
generationConfig: { temperature: 0.2, // Lower = more deterministic topP: 0.8, // Nucleus sampling topK: 40, // Top-k sampling maxOutputTokens: 2048, // Increase for longer responses responseMimeType: "application/json" }
// In AgentOrchestrator.ts private classifyTableContent(table: ExtractedTable): ClinicalDataType { const content = table.headers.join(' ').toLowerCase(); // Patient Demographics if (matchesPattern(content, ['age', 'sex', 'gender', 'patient', 'n ='])) return 'patient_demographics'; // Surgical Procedures if (matchesPattern(content, ['surgery', 'procedure', 'sdc', 'evd'])) return 'surgical_procedures'; // Outcomes Statistics if (matchesPattern(content, ['outcome', 'mortality', 'mrs', '%'])) return 'outcomes_statistics'; // Neuroimaging Data if (matchesPattern(content, ['volume', 'lesion', 'swelling', 'mm'])) return 'neuroimaging_data'; // Study Methodology if (matchesPattern(content, ['study', 'design', 'inclusion', 'criteria'])) return 'study_methodology'; return 'unknown'; }
// Add to classifyTableContent() if (matchesPattern(content, ['your', 'keywords', 'here'])) return 'your_new_type';
// In browser console const testTable = { id: "test-1", pageNum: 1, headers: ["Characteristic", "Value"], rows: [["Age", "59.2±12.3"], ["Sex", "60.7% male"]], extractionMethod: "test" }; const result = await MedicalAgentBridge.callAgent( 'PatientDataSpecialistAgent', testTable, 'table' ); console.log('Agent Result:', result);
// Load PDF and run await runFullAIPipeline(); // Check enhanced results const state = AppStateManager.getState(); console.log('Enhanced Tables:', state.extractedTables);
// Process multiple tables with same agent const results = await Promise.all( tables.map(table => MedicalAgentBridge.callAgent(agent, table, 'table')) );
// Cache agent results (add to AgentOrchestrator) private resultCache = new Map<string, AgentResult>(); private getCacheKey(agentName: string, data: any): string { return `${agentName}:${data.id}`; }
// Only process high-confidence geometric extractions if (table.structureConfidence > 0.9) { await enhanceWithAgents(table); }
Problem: Agents returning <0.85 confidence
Solutions:
Problem:
Failed to parse response as JSONSolutions:
responseMimeType: "application/json"Problem: Too many API calls
Solutions:
✅ DO:
❌ DON'T:
Last Updated: November 15, 2025 Maintainer: Clinical Extractor Team Status: Production Ready