Abstract / Introduction

Finding clear, trustworthy guidance when you're worried about your health shouldn't feel overwhelming. This project proposes an AI-powered smart health assistant that listens to how people actually describe their symptoms by text or optional voice, and translates that into likely conditions with strong, evidence-based next steps. Our goal is to reach around 80% top-3 accuracy on a curated symptom-condition dataset, while offering practical triage advice (self-care, see a GP, or seek urgent care) and recommending the right specialist for 90%+ of valid queries. To keep the experience transparent, each result comes with an easy-to-read explanation (via SHAP/LIME) for 95%+ of outputs. The system is designed to feel fast and dependable responding in about 2.5 seconds on typical text queries and maintaining 99% uptime during the demo period. And because accessibility matters, the full interface is available in English and Urdu, with accurate voice transcription (target 90%+ on clean audio).

Safety note: This assistant offers preliminary guidance only and is not a medical diagnosis. A clear disclaimer appears on every screen, and users are encouraged to consult qualified healthcare professionals for clinical concerns.

Functional Requirements:

1. User Roles

• Guest User: Submit symptoms, view results, and read basic guidance.
• Registered User (Patient profile): Guest features plus saved history, preferences/language, and PDF export.
• Admin/Clinician Reviewer (optional): Manage knowledge base, view anonymized analytics, approve content updates.

2. Core Use Cases

UC-01: Submit Symptoms (Text)

• Input: Free-text description; optional metadata (age, sex, duration, severity 1–5), risk flags (pregnancy, chronic disease).
• Process: NLP preprocessing → symptom entity extraction → feature vector → ML/DL classifier.
• Output: Top-k likely conditions (with confidence), triage level (self-care / see GP / urgent care), recommended specialist.

UC-02: Submit Symptoms (Voice) — optional

• Input: Microphone audio (8–15s).
• Process: Speech-to-Text → same pipeline as UC-01.
• Output: Same as UC-01.

UC-03: Personalized Guidance

• Input: User profile (age/sex/history) and current symptoms.
• Process: Rule-augmented post-processor (e.g., if fever + age <5 → urgent care).
• Output: Next steps, home-care tips, red-flag warnings, specialist type, and when to seek care.

UC-04: Explainability Panel

• Input: Model output and feature representation.
• Process: SHAP/LIME explainer on the final model.
• Output: Top contributing symptoms/phrases, feature-importance plot, and a short rationale sentence.

UC-05: Multilingual UI

• Process: UI strings via i18n; Urdu/English toggle with RTL support.
• Output: Full UI translated; model output labels localized.

UC-06: History & Export (Registered)

• Actions: Save queries and outputs; export last N results as PDF; delete history (GDPR-like).
• Output: Downloadable PDF with disclaimer and timestamp.

UC-07: Admin Knowledge Base (optional)

• Actions: Add/edit symptom synonyms, map conditions→specialist, edit self-care templates, review logs.
• Output: Versioned content with rollback.

UC-08: Feedback Loop

• Input: 1–5 helpfulness rating and free-text feedback.
• Output: Stored feedback; optional model-monitoring dashboard.

3. Data & Models

• Datasets: Public symptom–condition mappings (e.g., Kaggle/academic) plus curated synonym lists.
• Preprocessing: Tokenization; symptom entity extraction; spelling normalization; optional Urdu transliteration support.

Model Options:

• Baseline: Logistic Regression / SVM / RandomForest on multi-label bag-of-symptoms.
• NLP-Enhanced: Clinical-style embeddings (e.g., sentence-BERT) feeding a classifier head.
• Hybrid: Rule engine for red flags combined with ML predictions.
• Explainability: SHAP Tree/Kernel explainers for tabular; LIME/SHAP for text.

4. Inputs / Outputs & Validation

Inputs:

• Text: 5–500 characters; block PII (names/phone/address) with regex before storing.
• Voice: 16kHz mono WAV/WEBM; max 20s; transcribe then discard raw audio (configurable).

Outputs:

• Top-3 conditions with confidence (0–1), triage level, specialist type, explainability plot, guidance bullets, and disclaimer.

Validation & Errors:

• Empty/short text → prompt for more details.
• Unsafe content (self-harm) → show emergency resources message.
• Ambiguous symptoms → ask clarifying follow-ups (fever? duration?).

5. Workflows

• Symptom Flow: Input → Preprocess → Predict → Post-process (rules) → Explainability → Render → Save (if logged in).
• Admin Content Flow: Edit KB → Validate schema → Save draft → Preview → Publish → Version.

6. Security & Privacy

• Do not store raw audio by default; store only transcriptions if the user consents.
• Hash user IDs; encrypt at rest (DB-level).
• Role-based access control (RBAC) for Admin areas.
• Include Terms/Privacy pages; explicit medical disclaimer on every result screen.