Week 2: AI Architecture and Algorithms¶

Section 1: Foundations | Learning Outcome 1¶

Theme: Under the hood: How healthcare AI systems actually work

Core Question: "What is this?"

Learning Objectives¶

By the end of this session, students will be able to:

Explain the fundamental architecture of machine learning systems
Differentiate between supervised, unsupervised, and reinforcement learning
Describe how neural networks process healthcare data
Understand the basics of large language models and their healthcare applications

Content¶

2.1 Machine Learning Fundamentals¶

The Learning Paradigm

Unlike traditional software (where programmers write explicit rules), machine learning systems learn patterns from data:

Training: The system is exposed to many examples with known outcomes
Validation: Performance is tested on held-out data during development
Testing: Final evaluation on completely unseen data
Deployment: Application to new, real-world cases

Supervised Learning

The system learns from labelled examples—inputs paired with correct outputs.

Classification: Predicting categories - Is this lesion malignant or benign? - Which patients will develop sepsis? - What diagnosis does this presentation suggest?

Regression: Predicting continuous values - What is this patient's risk score (0-100)? - How long will this patient stay in hospital? - What is the predicted blood glucose in 2 hours?

Unsupervised Learning

The system finds patterns in unlabelled data.

Clustering: Grouping similar cases - Patient phenotyping (identifying natural subgroups) - Disease subtype discovery - Healthcare utilisation patterns

Dimensionality reduction: Simplifying complex data - Visualising high-dimensional patient data - Feature extraction from images or signals

Anomaly detection: Identifying outliers - Unusual lab patterns suggesting error or rare conditions - Fraud detection in claims data

Reinforcement Learning

The system learns through trial and error, receiving rewards for good outcomes.

Treatment optimisation (adjusting therapy based on response)
Resource allocation (learning efficient scheduling)
Less common in clinical practice due to safety concerns

2.2 Neural Networks and Deep Learning¶

From Perceptrons to Deep Networks

Neural networks are inspired by (but don't truly replicate) biological neurons: - Input layer: Receives data (vital signs, pixel values, text) - Hidden layers: Transform and combine information - Output layer: Produces predictions

"Deep learning" simply means neural networks with many hidden layers, enabling them to learn complex patterns.

Convolutional Neural Networks (CNNs) for Medical Imaging

Specialised architecture for image analysis: - Automatically learns visual features (edges, textures, shapes) - Hierarchical learning: simple features → complex patterns - Applications: radiology, pathology, dermatology, ophthalmology

Recurrent Networks and Transformers for Sequential Data

For data with temporal or sequential structure: - Vital sign trends over time - Clinical note sequences - Patient trajectories

The "Black Box" Problem

Deep learning models are notoriously difficult to interpret: - Millions of parameters with no clear meaning - Cannot easily explain why a prediction was made - Tension between predictive power and transparency - Implications for clinical trust and accountability

2.3 Large Language Models in Healthcare¶

How LLMs Work

Large Language Models (like GPT-4, Claude, Med-PaLM) are a specific type of neural network trained on vast text data:

Tokenisation: Text is broken into units (words or word-pieces)
Attention: The model learns which parts of the input relate to each other
Generation: Produces output one token at a time, predicting what comes next

Clinical Applications

Documentation assistance and summarisation
Clinical question answering
Differential diagnosis brainstorming
Patient communication drafting
Literature synthesis

Critical Limitations

Hallucination: Generating plausible but false information
Training data biases: Reflecting biases in source material
Lack of grounding: No connection to real-time patient data or truth
Overconfidence: Presenting uncertain information with confidence
Knowledge cutoffs: Training data has a fixed endpoint

Current Landscape

General models (GPT-4, Claude) with medical capability
Medically fine-tuned models (Med-PaLM, various clinical LLMs)
Integrated clinical systems with LLM components
Rapidly evolving—anything written today may be outdated soon

Aeromedical Thread¶

Architecture Choices for Retrieval Challenges¶

Different AI architectures suit different retrieval medicine applications:

Challenge	Suitable Architecture	Example
Point-of-care ultrasound interpretation	CNN (image classification)	"Is there pericardial effusion?"
In-flight deterioration prediction	Recurrent/time-series models	Early warning from continuous monitoring
Handover documentation	LLM (text generation)	Structured summary from raw notes
Triage prioritisation	Classification models	Dispatch urgency scoring
Resource allocation	Reinforcement learning or optimisation	Which asset for which case?

Connectivity Considerations

Which models can run locally (edge computing) vs. require cloud connectivity?
Latency requirements for real-time applications
Graceful degradation when connectivity fails

Learning Activities¶

Pre-Class Preparation¶

Interactive ML Module
Complete the provided online module on ML fundamentals (approx. 45 mins)
Take notes on concepts you find confusing
LLM Experimentation
Using a public LLM (ChatGPT, Claude, etc.), test it with clinical scenarios
Document what it does well and where it fails

In-Class Activities¶

Neural Network Walkthrough (Facilitated, 30 mins)
Interactive visualisation of a simple neural network
Trace how inputs become outputs
Adjust parameters and observe changes
Architecture Matching Exercise (Small Groups, 20 mins)
Groups receive clinical problems
Task: Identify which AI architecture would be appropriate and why
Present reasoning to class
LLM Behaviour Exploration (Hands-on, 30 mins)
Structured prompts to test LLM capabilities and limitations
Compare responses across different prompt formulations
Identify patterns in errors and hallucinations

Post-Class Activities¶

Technical Summary Exercise
Choose one AI architecture covered this week
Write an explanation (500 words) suitable for a clinical colleague
Peer review in pairs before next session

Practical Exercise 2: ML Fundamentals Hands-on (Colab)¶

📓 Open Exercise 2 Notebook

Objective¶

Train a simple classifier to understand the ML process from data to prediction.

Exercise Overview¶

This exercise demystifies machine learning by having you train a model yourself using a simplified clinical dataset.

Part A: Preparing the Data¶

# You'll work with a dataset of patient observations
# Goal: Predict which patients will deteriorate within 24 hours
# Features: vital signs, lab values, demographics
# Label: deteriorated (yes/no)

import pandas as pd
data = pd.read_csv('sample_deterioration_data.csv')

print(data.head())
print(f"Dataset has {len(data)} patients")
print(f"Deterioration rate: {data['deteriorated'].mean():.1%}")

Part B: Training Your First Model¶

from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier

# Split data into training and testing sets
# This is crucial—we test on data the model hasn't seen
X_train, X_test, y_train, y_test = train_test_split(
    features, labels, test_size=0.2, random_state=42
)

# Train a simple decision tree classifier
model = DecisionTreeClassifier(max_depth=5)
model.fit(X_train, y_train)

# See how well it performs
accuracy = model.score(X_test, y_test)
print(f"Model accuracy: {accuracy:.1%}")

Part C: Understanding Model Behaviour¶

Examine which features the model considers important
Visualise the decision tree to see the "reasoning"
Test what happens when you change training parameters
Observe how performance changes with different train/test splits

Part D: Experimentation Tasks¶

Modify the notebook to explore:

What happens if you train on only 50 patients vs. 500?
What if the training data has a different deterioration rate than reality?
What if you remove vital signs and only use demographics?

Key Learning Points¶

ML models learn patterns from training data—they can't know what they weren't shown
Test performance may not match training performance (overfitting)
The features you include determine what the model can learn
Random variation affects results—the same data can produce different models

Deliverable¶

Complete the guided notebook; answer embedded reflection questions.

Indicative Resources¶

Required Reading¶

Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine Learning in Medicine. New England Journal of Medicine, 380(14), 1347-1358.

Interactive Resources¶

TensorFlow Playground: playground.tensorflow.org
CNN Explainer: Visual tool for understanding convolutional neural networks

Session Summary¶

This week provided technical foundations for understanding how AI systems work:

Machine learning systems learn patterns from data through training
Supervised, unsupervised, and reinforcement learning address different problems
Neural networks power modern AI, with architectures specialised for different data types
Large language models represent a step-change in AI capabilities—and new risks
The "black box" problem creates tension between performance and explainability

Next Week: We'll examine the data that powers AI systems—how it shapes behaviour, creates biases, and determines whether systems generalise to new contexts.