Onco-Shikshak: An AI-Native Adaptive Learning Ecosystem for Medical Oncology Education

Onco-Shikshak V7 is a pioneering AI-native adaptive learning platform for medical oncology that mitigates LLM risks and addresses knowledge obsolescence by integrating a unified cognitive architecture of learning science principles, multi-disciplinary specialist agents, and closed-loop feedback mechanisms into authentic clinical workflows.

The Gist

Imagine trying to learn how to be a master chef in a world where new recipes are invented every single day, and the ingredients list changes every week. That is what studying Medical Oncology (cancer treatment) feels like today. The knowledge moves so fast that old textbooks become outdated before you even finish reading them.

Now, imagine you have a super-smart AI assistant to help you learn. But here's the catch: sometimes these AI assistants make things up (hallucinations), and sometimes they are too helpful, making you lazy so you stop thinking for yourself.

Onco-Shikshak is a new, smart learning system designed to solve these problems. Think of it not as a boring textbook, but as a "Cognitive Exoskeleton"—a high-tech suit that helps your brain lift heavy medical knowledge without getting crushed.

Here is how it works, explained through simple analogies:

1. The "Real-Life" Classroom (Situated Cognition)

Most learning apps are like flashcards: "What is the capital of France?" Click. "Paris."
Onco-Shikshak is different. It doesn't ask you to memorize facts in a vacuum. Instead, it drops you into four realistic scenarios where doctors actually work:

• Morning Report: You present a patient case to a senior doctor.
• Tumor Board: You sit in a meeting with six different specialists arguing about the best treatment.
• Clinic Day: You manage a busy waiting room with multiple patients.
• AI Textbook: You read a chapter that updates itself in real-time.

The Analogy: Instead of watching a video of someone swimming, you are thrown into the pool with a lifeguard. You learn by doing, not just by reading.

2. The "Smart Coach" (Adaptive Learning)

Imagine a personal trainer who knows exactly how strong you are.

• If you are a beginner, the trainer gives you heavy weights and holds your hand (Scaffolding).
• If you are an expert, the trainer takes away the hand-holding and throws you curveballs to keep you sharp.

Onco-Shikshak uses a math system called Item Response Theory to act as this trainer. It constantly guesses your skill level. If you get a question right, it makes the next one slightly harder. If you get it wrong, it gives you a hint. It never lets you get bored (too easy) or frustrated (too hard). It keeps you in the "Goldilocks Zone" of learning.

3. The "Six-Headed Brain" (Multi-Agent Deliberation)

This is the coolest part. In a real hospital, a cancer patient isn't treated by one person; they are treated by a team: a surgeon, a radiologist, a pathologist, etc.
Onco-Shikshak simulates this with six AI agents, each acting like a different specialist:

• Dr. Pathology looks at the tissue.
• Dr. Radiology looks at the X-rays.
• Dr. Surgery wants to cut it out.
• Dr. Radiation wants to zap it.
• Dr. Medicine wants to use drugs.
• Nurse Rivera worries about the patient's comfort and side effects.

The Twist: These AI agents often disagree. The surgeon might say, "Cut it out!" while the radiation doctor says, "No, let's use radiation."
The Lesson: You, the learner, have to listen to all of them, weigh the evidence, and make the final decision. This forces you to think critically instead of just accepting one answer. It's like being the referee in a debate between six experts.

4. The "Anti-Forgetting Machine" (Spaced Repetition)

We all know the feeling of studying for a test and forgetting everything two days later.
Onco-Shikshak uses a system called FSRS (Free Spaced Repetition Scheduler). Think of it as a smart calendar that knows exactly when you are about to forget something.

• It shows you a flashcard right before your memory fades.
• If you remember it, it waits longer before showing it again.
• If you forget it, it shows it to you sooner.

It also mixes things up (interleaving). Instead of studying only "Lung Cancer" for an hour, it might ask you a question about Breast Cancer, then switch back to Lung Cancer. This "desirable difficulty" makes your brain work harder, which makes the memory stickier.

5. The "Truth Detector" (No Fake News)

One of the biggest fears with AI is that it lies. Onco-Shikshak solves this by tethering every AI answer to nine real, authoritative medical guidelines (like the NCCN or ESMO).
Before the AI speaks, it checks its sources. If it suggests a drug, it must show you the specific page in the guideline that says it's okay. It's like a student who has to cite their sources on every single sentence they write.

6. The "Mirror" (Metacognition)

The system doesn't just ask, "Did you get it right?" It asks, "How sure were you?"

• If you say "I'm 100% sure" but get it wrong, the system gently points out, "You were overconfident."
• If you say "I'm guessing" but get it right, it says, "You got lucky, but let's make sure you actually know this."

This trains you to be honest with yourself about what you know and what you don't, which is the mark of a true expert.

The Big Picture

Onco-Shikshak is a closed-loop learning ecosystem.

1. You make a mistake in a case.
2. The system instantly turns that mistake into a flashcard for you to review later.
3. If you keep missing a certain topic, it suggests a new case to practice that specific skill.
4. Every interaction updates your "learner profile," so the system knows exactly how to help you next time.

In short: It's a video game where the goal isn't to get a high score, but to become a better doctor. It combines the best of human teaching (real-world scenarios, expert debates) with the best of AI (instant feedback, infinite patience, perfect memory), all while making sure you never stop thinking for yourself.

Technical Summary

1. Problem Statement

Medical oncology education faces a dual crisis:

1. Knowledge Velocity: The field evolves rapidly (knowledge half-life ~3.5 years), with guidelines (e.g., NCCN) updating multiple times a year. Static curricula and traditional question banks cannot keep pace.
2. LLM Risks: While Large Language Models (LLMs) offer generative capabilities, they introduce hallucinations (fabricating clinical facts) and automation bias (learners uncritically accepting AI answers, bypassing the cognitive effort required for expertise).

Existing AI tools often function as simple information retrieval or diagnostic aids, lacking the pedagogical scaffolding, adaptive difficulty, and spaced repetition mechanisms necessary for sustained, deep learning.

2. Methodology: The Onco-Shikshak V7 Architecture

The authors propose Onco-Shikshak V7, an AI-native platform designed as a "clinical brain exoskeleton." It integrates five core learning science frameworks into a unified cognitive architecture:

A. Core Cognitive Frameworks

• ACT-R (Adaptive Control of Thought—Rational): Implements "illness scripts" as data structures. It models memory using Base-Level Activation (based on retrieval frequency and recency) and Spreading Activation (linking related cancer types, e.g., NSCLC-EGFR to NSCLC-ALK).
• Item Response Theory (IRT): Uses the Rasch model to estimate a learner's latent ability ($\theta$) and item difficulty ($\beta$). This drives Adaptive Scaffolding, adjusting the difficulty of clinical reasoning tasks to keep the learner within their Zone of Proximal Development (ZPD).
• FSRS (Free Spaced Repetition Scheduler): Replaces older SM-2 algorithms with FSRS v4 to optimize flashcard scheduling based on stability, difficulty, and retrievability, ensuring long-term retention.
• Metacognitive Calibration: Trains learners to monitor their own confidence using Brier scores to measure the alignment between confidence ratings and actual accuracy.
• Situated Cognition: Moves away from isolated modules to authentic clinical workflows (Morning Report, Tumor Board, Clinic Day, AI Textbook).

B. System Architecture & Tech Stack

• Frontend/Backend: Built on Next.js 16 and React 19 with Prisma ORM (PostgreSQL/SQLite).
• AI Engine: Powered by Gemini 2.5-Flash.
• Multi-Agent Deliberation: Features six specialist AI agents (Medical Oncology, Radiation Oncology, Surgical Oncology, Pathology, Radiology, and Oncology Navigation).
  • Per-Agent RAG: Each agent queries a specific subset of nine authoritative guideline sources (e.g., Pathology agents query CAP/ClinVar; Radiation agents query ASTRO). This prevents cross-contamination of guidelines and grounds responses in evidence.
• Data Layer: 27 Prisma models managing authentication, learner modeling, session management, and analytics.

C. Learning Workflows

1. Morning Report: Progressive data revelation (lab/imaging hidden until ordered) with six-phase clinical reasoning scaffolding.
2. Tumor Board: A multi-agent simulation where specialists express genuine clinical disagreements (e.g., Surgery vs. Radiation), forcing the learner to synthesize evidence and adjudicate.
3. Clinic Day: Manages multiple patients with random interruptions to test prioritization and SOAP note documentation.
4. AI Textbook: Dynamic, RAG-grounded chapter generation with inline citations and linked clinical cases.
5. Spaced Repetition 2.0: Interleaved flashcard review with contextual variants.

3. Key Contributions

1. Unified Cognitive Architecture: The first system to integrate ACT-R, IRT, FSRS, ZPD, and metacognitive calibration into a single medical education platform.
2. Multi-Agent RAG Deliberation: A novel implementation where six AI specialists deliberate with distinct, guideline-grounded perspectives, creating "cognitive conflict" to force higher-order reasoning.
3. Closed-Loop Feedback System: Four automated feedback loops:
   • Error $\to$ Card: Session errors generate targeted flashcards.
   • Weakness $\to$ Case: Repeated errors in a domain trigger suggested clinical cases.
   • Textbook $\to$ Case: Reading chapters links to relevant practice cases.
   • Interaction $\to$ Profile: All data updates a persistent learner model.
4. Competency Mapping: Every interaction is mapped to 13 ACGME Hematology-Oncology milestones and 5 Entrustable Professional Activities (EPAs), providing portfolio-ready evidence.
5. Guideline-Concordant Case Library: 18 clinical cases across 6 cancer types (NSCLC, Breast, Colorectal, Prostate, Lymphoma, Leukemia) with decision trees based on NCCN Category 1–2A evidence.

4. Results & Technical Validation

The paper reports technical validation of eight system properties (no formal learner RCT has been conducted yet):

• RAG Grounding: Verified that citations are traceable to source documents (NCCN, ESMO, etc.).
• FSRS Correctness: Stability and retrievability curves match the reference specification; migration from SM-2 preserves continuity.
• IRT Convergence: Learner ability estimates ($\theta$) converge within $\pm0.5$ standard error after ~20 interactions.
• ACT-R Dynamics: Base-level activation and spreading activation function correctly according to the logarithmic summation formula.
• Agent Isolation: Confirmed that agents only access their designated guideline sources (e.g., Pathology agents do not see Radiation guidelines).
• Metacognition: Brier score calculations and error pattern recognition function as designed.
• ACGME Mapping: All interaction types correctly map to milestone codes.

5. Significance

• Paradigm Shift: Moves medical education from static information delivery to dynamic clinical apprenticeship.
• Mitigating AI Risks: By introducing "desirable friction" (Socratic questioning, confidence elicitation) and grounding AI in specific guidelines, the system combats hallucination and automation bias.
• Scalable Expertise: The system operationalizes the transition from novice to expert by adapting scaffolding levels (Full $\to$ Partial $\to$ Minimal $\to$ Challenge) based on real-time psychometric data.
• Open Source: The code, case library, and schema are available under the MIT license, promoting reproducibility and community development.

Limitations & Future Work:
The authors acknowledge the lack of formal learner evaluation (a Randomized Controlled Trial is planned), uncalibrated IRT parameters for specific oncology populations, and limited cancer type coverage. Future work includes longitudinal case chains, semantic search for local textbooks, and LMS integration.