AI in Health Care (AL-504 (A)) - Unit 2 Short Notes

UNIT 2: AI in Healthcare - Short Notes

I. Foundations and Evaluation

Overview of AI Applications in Healthcare

AI is transforming healthcare across four broad domains:

1. Diagnostics: Analyzing medical images (X-rays, MRIs), pathology slides, and genomic data for early disease detection.

2. Treatment: Personalizing treatment plans, drug discovery, robotic surgery assistance, and predicting treatment response.

3. Administration: Automating billing, scheduling, claims processing, and clinical documentation.

4. Patient Engagement: Powering chatbots, virtual health assistants, and personalized health recommendations.

Exam Tip: Be prepared to cite specific examples for each domain (e.g., AI in radiology for diagnostics, chatbots for patient engagement).

Evaluation Metrics for AI Models

Selecting the right metric is critical due to the high-stakes nature of healthcare.

Common Pitfall: Do not rely solely on Accuracy for imbalanced datasets (e.g., 99% healthy patients). A model predicting "all healthy" would be 99% accurate but useless.

II. Remote Monitoring and Wearable Technology

Remote Patient Monitoring with AI

Architecture & Data Flow:

1. Sensing: Wearable/implantable sensors (ECG, SpO2, glucose) collect continuous physiological data.

2. Transmission: Data sent via Bluetooth/Wi-Fi to a gateway (smartphone) and then to cloud servers.

3. AI Analysis: Real-time stream processing (using models like LSTMs) detects anomalies (e.g., arrhythmia), predicts decompensation, and generates alerts.

4. Clinical Action: Alerts are routed to clinicians or caregivers via dashboards/mobile apps.

Key Benefit: Enables proactive intervention, reduces hospital readmissions, and supports chronic disease management (e.g., CHF, diabetes).

Wearable Health Technology and AI

Device Types: Smartwatches (ECG, activity), patches (continuous glucose), smart rings (sleep, temperature), implantables (loop recorders). Data Integration: AI fuses multi-modal data (physiological, activity, sleep) with EHR data for a holistic patient view. Clinical Decision Support: AI models on wearable data can:

• Predict atrial fibrillation from PPG signals.

• Estimate blood pressure from pulse wave analysis.

• Detect early signs of infection via elevated resting heart rate & temperature.

Challenge: Data noise, motion artifacts, and ensuring algorithmic robustness across diverse populations.

III. Hospital Operations and Resource Management

AI for Hospital Resource Optimization

• Staffing: Predictive models forecast patient admission rates (using historical data, seasonality, local events) to optimize nurse/doctor schedules.

• Bed Management: Reinforcement learning models predict patient discharge times and admission bottlenecks to maximize bed turnover.

• Inventory & Supply Chain: Time-series forecasting (e.g., ARIMA, Prophet) predicts demand for medicines, PPE, and surgical supplies, automating reordering.

• Scheduling: Optimization algorithms (Integer Programming) schedule surgeries, equipment, and staff to minimize idle time and wait times.

AI in Emergency Room Triage

• Prioritization Algorithms: ML models (e.g., gradient boosting) predict patient acuity and risk of critical events (e.g., sepsis, cardiac arrest) using initial vitals, chief complaint, and lab results.

• Workflow Integration: AI tools provide a risk score to triage nurses, augmenting (not replacing) clinical judgment. They can flag high-risk patients in waiting rooms, reducing time-to-treatment for life-threatening conditions.

• Example: The ESI (Emergency Severity Index) can be enhanced with AI predictions for more dynamic prioritization.

IV. Prognostic and Predictive Modeling

Linear Prognostic Models

• Concept: Use linear regression (or logistic regression for binary outcomes) to predict a future health outcome (e.g., 5-year survival, readmission risk) based on current predictors (age, biomarkers, comorbidities).

• Formulation (Logistic for Binary Outcome):

$$ \log\left(\frac{p}{1-p}\right) = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + ... + \beta_k X_k $$

where \( p \) is the probability of the event (e.g., death), \( X_i \) are predictors (e.g., age, cholesterol), and \( \beta_i \) are coefficients.

• Medical Interpretation: \( \beta_i \) represents the log-odds ratio per unit increase in \( X_i \), holding others constant. Easily interpretable for clinicians (e.g., "each 10 mmHg increase in systolic BP increases log-odds of stroke by 0.3").

Survival Analysis in Healthcare

Purpose: Analyze time-to-event data (e.g., time until death, relapse, device failure), handling censoring (patients lost to follow-up or event not occurred by study end).

Survival Models vs. Time-Dependent Survival Models

Nelson-Aalen Estimator

• Purpose: Non-parametric estimate of the cumulative hazard function \( H(t) \).

• Formula:

$$ \hat{H}(t) = \sum_{i: t_i \leq t} \frac{d_i}{n_i} $$

where at each distinct failure time \( t_i \), \( d_i \) = number of events, \( n_i \) = number at risk *just before* \( t_i \).

• Application: Provides a raw estimate of the total risk accumulated over time. The survival function can be derived: \( \hat{S}(t) = \exp(-\hat{H}(t)) \).

• Diagram:

  DiagramCANVAS: Step-wise plot of Nelson-Aalen cumulative hazard estimate over time, with jumps at each event time proportional to d_i/n_i.

Survival Trees

• Construction: Similar to CART, but splits are based on log-rank test or weighted log-rank to maximize difference in survival distributions between child nodes.

• Splitting Criteria: Finds variable and cutoff that create two groups with the most statistically significant separation in their Kaplan-Meier curves.

• Output: A tree where each terminal node (leaf) provides a survival curve or median survival time for patients in that subgroup.

• Clinical Example: A tree might first split on "Tumor Stage > II", then on "Age > 65", defining distinct prognostic subgroups with different survival outcomes.

• Advantage: Captures non-linear interactions and is highly interpretable as a clinical decision rule.

Conditional Average Treatment Effect (CATE)

• Definition: The average causal effect of a treatment (T) on an outcome (Y) for individuals with specific covariates (X=x). It answers: "What is the treatment effect for a patient like this?"

$$ \tau(x) = E[Y(1) - Y(0) | X=x] $$

where \( Y(1) \) and \( Y(0) \) are potential outcomes under treatment and control.

• Why Used in Personalized Medicine: Moves beyond estimating the Average Treatment Effect (ATE) for a population. CATE identifies which subpopulations benefit most (or are harmed) by a treatment.

• Estimation Methods:

  1. Meta-learners: Algorithms like T-learner (separate models for treated/control), S-learner (single model with treatment as feature), and X-learner (combines both).

  2. Causal Forests: An extension of random forests that estimates CATE across the feature space.

• Example: Estimating the effect of a new drug on blood pressure reduction specifically for patients with a certain genetic marker and comorbidity profile.

Heart Disease Prediction

• Feature Engineering: Combines traditional risk factors (age, sex, cholesterol, BP, smoking) with derived features (e.g., BMI, cholesterol/HDL ratio, Framingham risk score components).

• Model Selection: Common models include Logistic Regression (interpretable baseline), Random Forests (handles non-linearity, feature importance), Gradient Boosting (XGBoost, often top performance), and Neural Networks.

• Risk Stratification: Output is typically a probability score (0-1). Patients are stratified into risk categories (Low/Medium/High) based on thresholds, guiding preventive interventions (lifestyle, statins, further testing).

• Key Challenge: Model calibration is vital—predicted probabilities must match observed event rates across all risk deciles.

V. Medical Imaging and Segmentation

AI in Medical Image Segmentation

• Goal: Pixel/voxel-level classification to delineate structures (organs, tumors, lesions) in 2D/3D medical images (CT, MRI, ultrasound).

• Key Technique: U-Net Architecture.

  • Encoder (Contracting Path): Standard CNN (convolutions + pooling) captures context and extracts features.

  • Decoder (Expansive Path): Transposed convolutions (upsampling) recover spatial resolution.

  • Skip Connections: Concatenate encoder feature maps with decoder layers to preserve fine-grained spatial details lost during pooling. This is the core innovation enabling precise localization.

  • DiagramCANVAS: U-shaped diagram showing the encoder (left, down-sampling), bottleneck, and decoder (right, up-sampling) with horizontal arrows (skip connections) linking corresponding layers.

• Applications: Tumor volume measurement (oncology), organ atlas creation (radiology), surgical planning, radiotherapy target delineation.

• Challenges: Limited annotated data, class imbalance (small lesion vs. large background), domain shift (scanner variations), and need for 3D segmentation (computationally intensive).

VI. Electronic Health Records (EHR) Systems

AI for EHR Efficiency

• Data Extraction & Structuring: NLP (e.g., BERT-based models) extracts structured information from unstructured clinical notes (e.g., symptoms, diagnoses, social history) to populate structured fields.

• Clinical Decision Support (CDS): AI analyzes patient's EHR in real-time to:

  • Flag drug-drug interactions or allergies at prescription time.

  • Suggest potential diagnoses based on symptoms and lab results (diagnostic support).

  • Generate automated care pathways or order sets.

• Workflow Automation:

  • Auto-coding: Automatically assign billing/ICD-10 codes from clinical documentation.

  • Smart Phrases: NLP-powered templates that auto-complete repetitive documentation.

  • Prioritization: AI sorts inbox messages (test results, consults) by urgency for clinicians.

VII. Specialized Healthcare Applications

AI in Disease Detection

• Early Diagnosis: AI models detect subtle patterns in imaging (e.g., micro-calcifications in mammograms), genomics, or longitudinal EHR data before clinical symptoms manifest.

• Screening Tools: Deployed as second readers (e.g., diabetic retinopathy screening from fundus photos) to increase throughput and reduce human error.

• Biomarker Identification: Unsupervised learning (clustering) or deep learning on multi-omics data (genomic, proteomic) discovers novel biomarkers for disease subtypes or prognosis.

Virtual Health Assistants (VHAs)

• Conversational AI: Powered by NLP (intent recognition, dialogue management, NLG). Can be rule-based or LLM-based.

• Patient Interaction: Symptom checking (triage), medication reminders, answering FAQs about conditions/treatments, mental health support (e.g., Woebot).

• Chronic Disease Management: Daily check-ins, data logging (glucose, BP), personalized education, and escalating alerts to care teams if parameters worsen.

Medication Adherence

• Monitoring Systems: Smart pill bottles (track openings), ingestible sensors, pharmacy refill data analysis.

• Predictive Alerts: ML models predict non-adherence risk based on patient history, social determinants, and medication complexity.

• Intervention Strategies: AI triggers tailored interventions: automated SMS/phone reminders, pharmacist outreach, or adjusting medication regimens for high-risk patients.

Sentiment Analysis in Healthcare

• Patient Feedback: Analyzing reviews, surveys, and social media to gauge patient satisfaction, identify service issues (long wait times, rude staff), and improve experience.

• Mental Health Assessment: Detecting depressive or anxious language from patient journal entries or therapy transcripts (as a supplementary tool).

• Service Improvement: Aggregating sentiment trends over time to evaluate the impact of hospital initiatives or staff training.

VIII. Machine Learning Fundamentals for Healthcare

Overfitting in Healthcare Models

• Definition: Model learns noise/irrelevant patterns in the training data, leading to poor generalization to unseen patient data. Performance drops on new data.

• Causes in Healthcare:

  • Small sample size relative to features (high-dimensional data like genomics).

  • Noisy, messy real-world EHR data.

  • Complex models (deep neural networks) with too many parameters.

• Detection:

  • Large gap between training accuracy (very high) and validation/test accuracy (significantly lower).

  • Performance degrades on external validation datasets from different hospitals.

• Mitigation Techniques:

  1. Regularization: Add penalty to loss function (L1/Lasso for sparsity, L2/Ridge for small weights).

  2. Cross-Validation (CV): Use k-fold CV to get robust performance estimate and tune hyperparameters. Stratified CV is crucial for imbalanced outcomes.

  3. Feature Selection/Dimensionality Reduction: Remove irrelevant features (using domain knowledge or statistical tests) or use PCA.

  4. Simplify Model: Prefer interpretable models (logistic regression) or reduce neural network depth/width.

  5. Early Stopping: For iterative models (gradient boosting, neural nets), stop training when validation error stops improving.

  6. Data Augmentation: Artificially increase training data size (e.g., for images: rotations, flips).

  7. Ensemble Methods: Bagging (e.g., Random Forest) reduces variance by averaging multiple models.

Exam Tip: Always link overfitting to the "validation on external/hold-out data" principle in healthcare. A model that overfits its training hospital data will fail in another hospital.