MODULE MAP

When the hospital's vendor NLP system nearly causes a fatal drug interaction, Dr. Karimi assembles a team to build Diagnostic One — an in-house clinical text AI. You'll learn why ML matters for text, how to convert words to numbers, and run your first text classifier.

Clinical text is messy — abbreviations, misspellings, shorthand. Imran presents the messiest patient records, and Vikram shows you how tokenizers break text into meaningful units, just like a URL router parses path segments.

Vikram demonstrates that 'aspirin' lives closer to 'ibuprofen' than to 'appendectomy' in vector space. You'll learn to represent words as dense vectors that capture semantic meaning — from naive one-hot encoding to pre-trained medical embeddings.

'Patient denied chest pain' versus 'Patient reported chest pain' — one word changes everything. You'll build models that understand sequence, from simple RNNs to bidirectional LSTMs, learning to extract structured clinical information from unstructured narratives.

When you read 'The patient was prescribed methotrexate for rheumatoid arthritis,' your brain automatically connects 'methotrexate' to 'arthritis.' You'll build this same selective focus into Diagnostic One — teaching the model which words matter most using attention mechanisms that work exactly like CSS :focus.

RNNs read text one word at a time. Transformers read everything at once — and they changed NLP forever. You'll build transformer blocks from scratch, understand BERT's bidirectional magic, and fine-tune a pre-trained model on clinical text. Along the way, MEDI makes a dangerous mistake: flagging a resolved condition as active.

Dr. Karimi needs structured information from unstructured clinical notes — drug names, conditions, dosages buried in free text. You'll build NER models that identify and classify medical entities, culminating in a breakthrough moment where your model catches a drug interaction that three doctors missed.

Hundreds of clinical documents arrive every hour. Which ones need immediate attention? You'll build classification and triage systems that route documents by urgency, assign multiple labels, and — critically — know when they don't know. Because in medicine, a confident wrong answer is more dangerous than an honest 'I'm not sure.'

An unconscious ER patient arrives with no accessible medical history. Records exist in a different system, partially in Spanish. 'Has this patient ever had an allergic reaction to penicillin?' You'll build extractive QA, reading comprehension, and RAG pipelines — because finding answers in documents is just React Server Components with a vector database.

Khalil leads the ethics review. 'What happens when it is wrong? Who is liable?' Dr. Karimi delivers the hard truth: the model performs 12% worse on records written in non-standard English. You'll build production-grade NLP APIs, optimize for clinical-grade latency, implement HIPAA-compliant privacy controls, audit for bias, and design human-in-the-loop approval gates.

Diagnostic One goes live. Day 9: MEDI flags a drug interaction between methotrexate and trimethoprim. Dr. Karimi confirms it is a false positive — but 'better a false alarm than a missed one.' You'll design the complete system architecture, integrate with hospital systems, validate against expert annotations, write clinical documentation, and deploy Diagnostic One to Signal Ward.