Feature Branch: 001-no-show-predictor
Created: 2026-01-28
Status: Draft
Input: User description: "Build a demonstration that includes a machine learning model and chat agent to show how an ML model output combined with Gen AI can surface likely no-show medical appointments and suggest actions."
Note: These are implementation-level decisions captured for context; see Requirements for functional scope.
| Component | Technology | Notes |
|---|---|---|
| ML Training | Azure AutoML | Automated feature engineering and model selection |
| ML Hosting | Azure ML Managed Online Endpoints | Real-time inference with managed scaling |
| Agent Runtime | Azure AI Foundry Hosted Agents | Containerized agent with conversation state management |
| Agent Framework | Microsoft Agent Framework (.NET) | Microsoft.Agents.AI 1.0.0-preview.260127.1 |
| Agent Language | .NET 10 (C#) | Latest preview with hosting adapter |
| Frontend | Blazor WebAssembly | Single-page app with dark/light mode, auto-scroll chat |
| Frontend Hosting | Azure Static Web Apps | Serverless hosting with global CDN |
| Infrastructure | Terraform AzApi Provider 2.8.0 | Native Azure ARM API access |
| Deployment | Azure Developer CLI (azd) | azd ai agent extension for hosted agents |
| Authentication | Managed Identity / DefaultAzureCredential | Credential-less auth per constitution |
| Region | North Central US | Required for hosted agents preview |
As a scheduling coordinator, I want to ask the system which patients are most likely to miss their appointments tomorrow so I can proactively reach out to confirm attendance.
Why this priority: This is the core value proposition—identifying no-show risks before they happen enables proactive intervention and reduces revenue loss from empty appointment slots.
Independent Test: Can be fully tested by asking "Which patients are at risk of missing their appointments tomorrow?" and receiving a ranked list of appointments with no-show probabilities and patient context.
Acceptance Scenarios:
-
Given appointments are scheduled for tomorrow, When the coordinator asks "Which patients are most likely to miss their appointments tomorrow?", Then the system returns a ranked list of appointments sorted by no-show probability with patient name, appointment time, provider, and risk score.
-
Given appointments exist with varying risk levels, When the coordinator asks "Show me high-risk appointments for tomorrow morning", Then the system filters to appointments before noon with no-show probability above a reasonable threshold (e.g., 50%).
-
Given no appointments are scheduled for tomorrow, When the coordinator asks about tomorrow's no-shows, Then the system responds that no appointments are scheduled for that date.
As a scheduling coordinator, I want the system to recommend specific actions I should take based on no-show predictions so I can efficiently manage my outreach and overbooking strategy.
Why this priority: Identifying risk is valuable, but actionable recommendations multiply the value by guiding staff on what to do about the risks.
Independent Test: Can be fully tested by asking "What scheduling actions should I take for tomorrow?" and receiving concrete recommendations like confirmation calls, overbooking suggestions, or reminder priorities.
Acceptance Scenarios:
-
Given high-risk appointments exist for tomorrow, When the coordinator asks "What actions should I take to reduce no-shows tomorrow?", Then the system provides prioritized recommendations such as "Call these 5 patients to confirm" or "Consider overbooking the 2pm slot with Dr. Smith."
-
Given a specific patient has a high no-show probability, When the coordinator asks "What should I do about the Johnson appointment?", Then the system explains the risk factors and suggests targeted interventions (e.g., "Patient has missed 3 of last 5 appointments—recommend personal phone call and offer transportation assistance if needed").
-
Given multiple providers have varying no-show patterns, When the coordinator asks "Which providers should I focus overbooking efforts on?", Then the system identifies providers with historically higher no-show rates and current high-risk appointment loads.
As a clinic manager, I want to see a summary of predicted no-shows for the upcoming week so I can plan staffing and resource allocation accordingly.
Why this priority: Weekly planning enables strategic decisions but is less time-sensitive than daily operational tasks.
Independent Test: Can be fully tested by asking "What does the no-show forecast look like for this week?" and receiving a daily breakdown of expected no-show counts and rates.
Acceptance Scenarios:
-
Given appointments are scheduled for the next 7 days, When the manager asks "What's the no-show outlook for this week?", Then the system provides a daily summary showing expected no-show count, percentage, and total appointments per day.
-
Given certain days have unusually high predicted no-show rates, When the manager asks about the weekly forecast, Then the system highlights concerning days and explains contributing factors (e.g., "Tuesday shows 35% predicted no-show rate—note: holiday weekend effect").
As a scheduling coordinator, I want to ask about a specific patient's no-show history and risk factors so I can have informed conversations when scheduling or confirming appointments.
Why this priority: Supports the other stories by providing drill-down capability, but is not essential for core no-show prevention workflow.
Independent Test: Can be fully tested by asking "Tell me about patient Maria Garcia's appointment history" and receiving a summary of past attendance, identified risk factors, and current risk assessment.
Acceptance Scenarios:
-
Given a patient has appointment history in the system, When the coordinator asks "What's the no-show history for patient ID 12345?", Then the system returns attendance statistics, identified risk factors, and the patient's current predicted no-show probability.
-
Given a patient is new with no history, When the coordinator asks about that patient, Then the system indicates limited history and explains which demographic or appointment factors contribute to the initial risk assessment.
- What happens when the ML model service is unavailable? System should gracefully degrade by indicating predictions are temporarily unavailable while still providing historical data.
- How does the system handle appointments with incomplete data (missing patient demographics)? System should still provide predictions using available features and note data quality limitations.
- What happens when a user asks about dates far in the future (e.g., 6 months out)? System should explain that predictions are most reliable within a shorter window and provide appropriate caveats.
- How does the system handle ambiguous date references (e.g., "next Tuesday" when today is Monday)? System should clarify the interpreted date in its response.
Synthetic Data Generation
- FR-001: System MUST generate synthetic patient appointment data following the provided schema with realistic distributions and patterns
- FR-002: Synthetic data MUST include historical no-show outcomes to enable model training and validation
- FR-003: Data generation MUST produce statistically valid patterns (e.g., higher no-show rates for certain demographics, appointment types, or time patterns based on published research)
Machine Learning Model
- FR-004: System MUST train a classification model to predict appointment no-show probability
- FR-005: Model MUST accept appointment and patient features as input and return a probability score between 0 and 1
- FR-006: Model MUST be deployable as a callable service endpoint
- FR-007: Model training MUST follow reproducible practices (fixed random seeds, versioned data, logged parameters)
- FR-008: Model MUST be validated using proper train/test split methodology to prevent data leakage
Chat Agent Interface
- FR-009: System MUST provide a conversational interface for querying no-show predictions
- FR-010: Agent MUST understand natural language queries about appointments, patients, and time periods
- FR-011: Agent MUST orchestrate data retrieval and model inference to answer user questions
- FR-012: Agent MUST provide actionable recommendations based on prediction results
- FR-013: Agent MUST explain risk factors contributing to high no-show predictions when asked
Data & Privacy
- FR-014: All patient data MUST be synthetic—no real patient information
- FR-015: System MUST NOT store conversation history containing patient identifiers beyond the current session
- The demonstration uses synthetic data only; no HIPAA compliance is required for this demo
- The chat interface is single-user (no concurrent user session management needed)
- Model retraining is out of scope; the demo uses a pre-trained model
- Integration with real EHR/scheduling systems is out of scope
- Patient: Individual with demographic information (age, gender, insurance status, distance from clinic), contact preferences, and appointment history
- Appointment: Scheduled visit with date/time, provider, appointment type, lead time (days scheduled in advance), and historical outcome (kept/no-show)
- Provider: Healthcare professional with specialty and schedule
- Prediction: Model output associating an appointment with a no-show probability and contributing risk factors
- Recommendation: Suggested action based on prediction (confirmation call, overbooking, reminder type)
- SC-001: Users can ask about no-show risks and receive accurate, relevant responses within 5 seconds
- SC-002: The ML model achieves at least 70% accuracy on held-out test data (demonstrating predictive value beyond random guessing)
- SC-003: 90% of natural language queries about appointments, patients, and scheduling actions are correctly interpreted and answered (measured against 20-query evaluation set covering date parsing, patient lookup, provider filtering, and action requests)
- SC-004: The system successfully generates at least 5,000 synthetic patient records and 100,000 appointment records spanning 2+ years to capture seasonality patterns
- SC-005: Recommendations provided by the agent are actionable and contextually appropriate (verified through sample query testing)
- SC-006: Model predictions include explainable risk factors that make clinical sense (e.g., "previous no-show history" as a top factor)