← Back to Portfolio
🤖 Agentic AI · LangChain · LangGraph · Claude Haiku · FastAPI

Patient Follow-Up
Agent

An agentic AI system that autonomously reviews patient records, flags clinical risk using defined medical thresholds, and generates prioritised care plans — built with LangChain 1.x, LangGraph orchestration, and Claude Haiku.

LangChain 1.x LangGraph Claude Haiku Agentic AI ReAct Pattern FastAPI Python Healthcare AI
Independent Project · Built and Validated Locally
100% tool call accuracy · 100% task completion · 100-patient synthetic dataset · 4 FastAPI REST endpoints tested
100%
Tool Call Sequence Accuracy
100%
Task Completion Rate
15
Missed Appointments Identified
8
High-Risk Patients Flagged
4
Custom Clinical Tools

🎯 Problem Statement

Clinical teams managing chronic disease patients at scale face a consistent challenge: manually reviewing records to identify who needs urgent follow-up is time-consuming, inconsistent, and reactive. High-risk patients slip through the cracks between appointments.

This project explores how an agentic AI system can automate that triage process — autonomously pulling patient records, assessing clinical risk against defined medical thresholds, generating prioritised care plans, and surfacing patients who have missed appointments. Built independently to deepen hands-on LangChain and agentic AI expertise, drawing on healthcare domain context from professional experience at Genpact.

🏗️ System Architecture

📋
CSV Dataset
100 patients
18 fields each
→
🔧
4 Custom Tools
load · assess
plan · missed
→
🧠
LangChain Agent
create_agent
LangGraph
→
🤖
Claude Haiku
Reasoning LLM
Anthropic API
→
⚡
FastAPI
4 REST endpoints
HTTP validated
ReAct Loop — Reason → Act → Observe → Repeat
Reason → Act (call tool) → Observe result → Reason again → Final response
Key design: The agent decides tool call order autonomously — no hardcoded sequence. Single-agent, tool-calling only. No RAG, no vector databases, no multi-agent orchestration.

🔧 4 Custom Clinical Tools

load_patient(patient_id)

Retrieves a single patient record from the dataset. Always called first before any analysis. Returns all 18 fields: vitals, lab values, diagnosis, medication, and appointment history.

assess_clinical_risk(patient_data)

Uses Claude Haiku to assess risk level HIGH / MEDIUM / LOW against defined clinical thresholds.

HbA1c > 9% → HIGH
BNP > 900 pg/mL → HIGH
BNP > 400 pg/mL → MEDIUM
BP > 160 mmHg → HIGH
eGFR < 30 → HIGH
PHQ-9 > 20 → HIGH
Days since visit > 300 → HIGH
generate_care_plan(patient_data)

Generates a prioritised follow-up care plan covering immediate actions, follow-up schedule, medication review, and patient communication. Capped at 150 words for clinical usability.

get_missed_appointments(days_threshold)

Returns all patients who missed their last appointment, filtered by minimum days since last visit. Deterministic data retrieval — no LLM call needed for this tool.

⚙️ Key Design Decisions

CHALLENGE
Token efficiency: sending full patient records to Claude Haiku for every tool call was expensive and slow at 100-patient scale.
SOLUTION
Compact key-value string format for patient records rather than JSON or prose — preserves all clinically relevant fields while reducing token count significantly.
CHALLENGE
Evaluation without ground truth: no clinician-labeled dataset existed to validate risk assessments using standard classification metrics.
SOLUTION
Process-level metrics — tool call sequence accuracy and task completion rate — plus output grounding spot-checks. A deliberate, defensible design choice given the prototype constraints.
CHALLENGE
Jupyter compatibility: running a FastAPI server in a notebook blocks further cell execution in the standard server startup pattern.
SOLUTION
Background thread pattern — server started in a daemon thread, cells continue running. Endpoints validated via direct HTTP client calls within the notebook.
CHALLENGE
Agent reliability: ensuring the LLM consistently calls tools in the correct clinical order without a hardcoded sequence.
SOLUTION
System prompt engineering — explicit workflow in the system prompt (load → assess → plan → summarise) combined with ReAct's observe-and-reason loop. Achieved 100% sequence accuracy across all test scenarios.

📊 Synthetic Dataset

100 synthetic patients across 8 chronic disease categories with 18 fields each. No real patient data was used at any stage.

100
Synthetic patients · 18 fields each
15
Patients with missed appointments
57
Patients overdue (>180 days)
211
Avg days since last visit
8
High-risk HF patients (BNP >900)
8
Disease categories covered

Diagnoses: Heart Failure, Type 2 Diabetes, COPD, Hypertension, CKD, Depression (PHQ-9), Hypothyroidism, Osteoarthritis.

⚡ FastAPI REST Endpoints

Agent exposed via 4 endpoints. Tested locally with server in a background thread, validated via direct HTTP client calls.

POST
/analyse-patient
Full analysis: load record → assess risk → generate care plan. Accepts patient_id.
GET
/missed-appointments
Returns all patients with missed appointments. Deterministic — no LLM call involved.
POST
/agent-query
Open-ended clinical query routed directly to the agent for ad-hoc requests.
GET
/health
System status: agent version, patient count, missed appointment count.

📈 Results and Evaluation

Evaluation used process-level metrics and output grounding spot-checks. No clinician-labeled ground truth was available, so classification metrics were not applicable. This is a deliberate, defensible design choice for a prototype of this nature.

Tool Call Sequence Accuracy
100% — agent called tools in the correct clinical order across all test scenarios
Task Completion Rate
100% — all queries returned structured, grounded clinical outputs
Missed Appointments
15 patients correctly identified by the get_missed_appointments tool
High-Risk Patients
8 heart failure patients flagged with BNP > 900 pg/mL threshold

💡 Key Technical Highlights

🔄 ReAct Agent Pattern

  • Agent reasons about which tool to call next based on prior observations
  • No hardcoded tool call sequence — fully autonomous orchestration
  • LangGraph manages the state graph under the hood via LangChain 1.x create_agent

🏥 Clinical Threshold Design

  • Risk thresholds based on established clinical guidelines: BNP, HbA1c, eGFR, PHQ-9
  • Structured prompts constrain LLM output to HIGH/MEDIUM/LOW + factors + urgency
  • Care plans capped at 150 words for real-world clinical usability

⚡ Token-Efficient Architecture

  • Compact key-value string format for patient records reduces LLM token cost
  • Deterministic tools bypass LLM entirely where no reasoning is needed
  • Claude Haiku chosen for cost efficiency at 100-patient testing scale

🧪 Defensible Evaluation

  • Process-level metrics used instead of classification metrics — no fabricated precision/recall
  • Output grounding spot-checks verify care plans reference actual patient data
  • Limitations clearly stated: synthetic data, no clinician validation, local prototype

🛠️ Technical Stack

🔗
LangChain 1.x
create_agent · Tool orchestration
📊
LangGraph
State graph · Agent orchestration
🤖
Claude Haiku
claude-haiku-4-5 · Reasoning LLM
⚡
FastAPI · Uvicorn
REST API · ASGI server
🐍
Python · Pandas
Data processing · Dataset management
📓
Jupyter Notebook
Development · Background thread testing
✅
Pydantic
Request/response validation
🔑
Anthropic API
Claude Haiku inference
⚠️ Note: Built independently to deepen hands-on agentic AI and LangChain expertise. Uses a 100-patient synthetic dataset only — no real patient data at any stage. Not a production deployment. Healthcare domain context drawn from professional experience at Genpact.