Onco-Shikshak

A retrieval-augmented AI platform for adaptive oncology education grounded in clinical practice guidelines

Medical oncology education demands continuous mastery of rapidly evolving clinical guidelines—the NCCN alone maintains over 76 guideline documents updated multiple times annually. We describe Onco-Shikshak, an AI-native adaptive learning ecosystem integrating three complementary modules: (1) a dynamic textbook generator producing expertise-level-adapted content grounded in retrieved guideline and textbook evidence, (2) a Socratic virtual preceptor guiding clinical reasoning through progressive disclosure and intentional cognitive friction, and (3) a spaced repetition engine generating atomic flashcards from authoritative sources scheduled via the SM-2 algorithm. The system employs retrieval-augmented generation (RAG) over seven clinical guideline corpora—NCCN, ESMO, ASTRO, ACR, CAP, ClinVar/CIViC, and SSO—with automated citation extraction from grounding metadata. The architectural design is grounded in six established learning science frameworks.

Three Integrated Learning Modalities

Each module implements specific learning science principles through concrete architectural decisions, sharing a common RAG knowledge grounding layer.

📖

Dynamic AI Textbook

Generates personalized oncology chapters on demand, adapted to Medical Student, Resident, or Attending expertise levels. Content is grounded in NCCN guidelines and foundational textbooks via RAG, with embedded retrieval practice checkpoints.

🧑‍⚕️

Virtual Preceptor

Dr. Chandra, an AI attending oncologist, uses Socratic questioning and progressive disclosure to guide clinical reasoning. Enforces NCCN compliance and introduces "desirable friction" to prevent automation bias.

💡

Spaced Repetition

Generates high-yield flashcards (MOA, TNM cutoffs, pathognomonic findings) from guideline and textbook context. SM-2 algorithm schedules reviews for optimal long-term retention.

Theory-to-Design Mapping

Every architectural decision maps to an established learning science framework.

Theory Principle Design Decision
Anderson (2000) Spacing effect & retrieval practice SM-2 algorithm with guideline cards
Yeo & Fazio (2019) Retrieval for facts; worked examples for procedures Flashcards + case simulations
Lee & Anderson (2013) Expertise reversal effect Three adaptive complexity levels
Tankelevitch (2025) Desirable friction; cognitive protection Socratic questioning in preceptor
Matuschak & Nielsen Learning in authentic context Clinical case immersion
LearnLM (2025) AI-augmented textbook RAG-grounded dynamic generation

7 Guideline Sources + 2 Textbook Corpora

All generated content is grounded in authoritative clinical sources via Gemini File Search RAG, with automated citation extraction from grounding metadata.

NCCN Guidelines ESMO Guidelines ASTRO Guidelines ACR Appropriateness Criteria CAP Cancer Protocols ClinVar / CIViC SSO Surgical Oncology DeVita 11th Ed. Abeloff 6th Ed.

Read the Paper

Full preprint describing the system design, architecture, pedagogical rationale, and technical validation.

Download PDF

Citation

@article{ashish2026oncoshikshak,
  title={Onco-Shikshak: A Retrieval-Augmented AI Platform
         for Adaptive Oncology Education Grounded in
         Clinical Practice Guidelines},
  author={Ashish},
  journal={arXiv preprint},
  year={2026},
  url={https://inventcures.github.io/onco-shikshak/}
}