AI in Health Care (AL-504 (A)) - Unit 3 Short Notes

UNIT 3: AI in Healthcare – Short Notes

I. General Applications Overview

AI applications in healthcare are broadly categorized based on their primary function and target user.

[!TIP] Exam questions often ask for specific examples within each category. Be ready to illustrate with real-world use cases (e.g., AI detecting diabetic retinopathy from retinal scans).

II. Diagnostic Applications

AI in Disease Detection

• Core Idea: Use ML/DL models to classify patient data (images, lab results, symptoms) as indicative of a specific disease.

• Process: Data (e.g., CT scan) → Pre-processing → Feature extraction (manual or via CNN) → Classification model (e.g., SVM, ResNet) → Disease label (Positive/Negative).

• Key Advantage: High throughput, consistency, and ability to detect subtle patterns beyond human perception.

• Example: Deep learning models (like U-Net) for pneumonia detection from chest X-rays with performance rivaling radiologists.

Medical Image Segmentation

• Definition: The pixel-level classification task of partitioning a medical image into meaningful regions (e.g., tumor vs. healthy tissue, organ boundaries).

• Primary Technique: Convolutional Neural Networks (CNNs), specifically encoder-decoder architectures like U-Net.

• Loss Function: Often uses Dice Loss or Jaccard Loss to handle class imbalance (small tumor vs. large background).

$$\text{Dice Coefficient} = \frac{2|X \cap Y|}{|X| + |Y|}$$

where $X$ = predicted segmentation, $Y$ = ground truth.

• Output: A binary mask outlining the region of interest (ROI).

Heart Disease Prediction (Diagnostic Case Study)

• Goal: Predict the presence of heart disease (e.g., atherosclerosis) from clinical and test data.

• Typical Data: Structured EHR data (age, cholesterol, BP, ECG results).

• Common Models: Logistic Regression (baseline), Random Forest, Gradient Boosting (XGBoost), or simple neural networks.

• Key Output: Probability score ($0$ to $1$) of having heart disease. Threshold (e.g., 0.5) determines final classification.

• Evaluation: Accuracy, Precision, Recall, F1-Score, AUC-ROC curve.

III. Prognostic and Risk Prediction Applications

Linear Prognostic Models

• Concept: Use linear regression or generalized linear models (GLMs) to predict a continuous future outcome (e.g., length of stay, cost) or a binary risk (e.g., 30-day readmission).

• Model: $$\displaystyle Y = \beta_0 + \beta_1 X_1 + ... + \beta_p X_p + \epsilon $$

  • $Y$: Prognostic outcome (e.g., survival time, risk score).

  • $$\displaystyle \beta_i $$: Coefficients indicating feature importance/direction.

• Advantage: Highly interpretable; clinicians can understand the impact of each variable ($$\displaystyle X_i $$).

• Limitation: Assumes linear relationships, may not capture complex interactions without feature engineering.

Survival Analysis

• Goal: Analyze time-to-event data, where the event is something like death, disease recurrence, or machine failure. Handles censored data (patients lost to follow-up or still alive at study end).

• Key Functions:

  1. Survival Function, $S(t)$: Probability of surviving beyond time $t$.

$$S(t) = P(T > t)$$

2.  **Hazard Function, $h(t)$:** Instantaneous event rate at time $t$, given survival up to $t$.

$$h(t) = \lim_{\Delta t \to 0} \frac{P(t \leq T < t+\Delta t | T \geq t)}{\Delta t}$$

• Common Models: Cox Proportional Hazards, Kaplan-Meier estimator (non-parametric).

Survival Model vs. Time Survival Model

• Survival Model (Cox PH): Models the hazard $h(t|X)$ as proportional to a baseline hazard $$\displaystyle h_0(t) $$.

$$h(t|X) = h_0(t) \exp(\beta^T X)$$

*   **Focus:** Effect of covariates ($X$) on the hazard.

*   **Assumption:** Hazard ratios are constant over time (proportionality).

• Time Survival Model (Parametric): Directly models the survival time $T$ with a specified distribution (Weibull, Exponential, Log-normal).

$$S(t|X) = S_0(t)^{\exp(\beta^T X)}$$

*   **Focus:** Explicitly models the shape of the survival curve over time.

*   **Use:** When proportionality assumption fails or for extrapolating survival beyond observed time.

Nelson-Aalen Estimator for Cumulative Hazard Function

• Purpose: Non-parametric estimate of the cumulative hazard function $$\displaystyle H(t) = \int_0^t h(u) du $$.

• Formula: At each distinct event time $$\displaystyle t_j $$:

$$\hat{H}(t) = \sum_{t_j \leq t} \frac{d_j}{n_j}$$

where $$\displaystyle d_j $$ = number of events at $$\displaystyle t_j $$, $$\displaystyle n_j $$ = number at risk *just before* $$\displaystyle t_j $$.

• Relation to Kaplan-Meier: $$\displaystyle \hat{S}(t) = \exp(-\hat{H}(t)) $$.

• Interpretation: Total "risk" accumulated up to time $t$.

Survival Tree

• Concept: A decision tree adapted for censored survival data.

• Splitting Criterion: Uses statistics like log-rank test or maximizing the difference in survival curves between child nodes, not just variance reduction.

• Output: Tree where each terminal node (leaf) provides a ** Kaplan-Meier survival curve** for the patients in that node.

• Advantage: Captures non-linear interactions and provides intuitive, rule-based risk stratification (e.g., "If Age > 60 & BP > 140, then high-risk group").

Conditional Average Treatment Effect (CATE)

• Definition: The individual-level estimate of the difference in outcome if a patient receives treatment A vs. treatment B.

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

where $Y(1)$/$Y(0)$ are potential outcomes under treatment/control, $X$ are patient features.

• Why Use It? Moves beyond average treatment effect (ATE) to personalized medicine. Answers: "What is the specific benefit/risk for this patient?"

• Estimation Methods: Meta-learners (e.g., T-learner, S-learner, X-learner) using ML base models (Random Forest, Gradient Boosting) to model outcome and treatment assignment.

IV. Healthcare Operations and Resource Management

Hospital Resource Optimization

• Problems: Staff scheduling, bed turnover, operating room (OR) scheduling, inventory management.

• AI Techniques:

  • Reinforcement Learning (RL): Learn optimal policies for dynamic resource allocation (e.g., bed assignment).

  • Integer Linear Programming (ILP) / Optimization: Formulate as cost-minimization or throughput-maximization problems with constraints.

  • Time Series Forecasting: Predict patient admission rates (using ARIMA, LSTM) to align staffing.

• Impact: Reduced wait times, lower operational costs, improved staff satisfaction.

Emergency Room (ER) Triage with AI

• Goal: Prioritize patients based on severity and predicted outcomes (e.g., risk of deterioration, length of stay).

• Input Data: Vital signs, chief complaint, nurse assessments, historical EHR data.

• Models: Gradient Boosting Machines (GBMs) or Random Forests trained to predict outcomes like "critical care need within 24h" or "hospital admission."

• How it Works: AI provides a risk score that supplements (not replaces) traditional triage (e.g., ESI). Flags high-risk patients for faster physician assessment.

• Benefit: Reduces mortality from missed deterioration, optimizes resource use in crowded ERs.

Electronic Health Record (EHR) System Efficiency

• Challenges: Data silos, unstructured notes, manual data entry, alert fatigue.

• AI Solutions:

  • Natural Language Processing (NLP): Extract structured data from clinical notes (e.g., symptoms, medications).

  • Predictive Coding: Suggest billing codes (ICD-10) based on documentation.

  • Smart Alerts: Reduce alert fatigue by using ML to prioritize only clinically relevant alerts (e.g., drug-drug interaction with high probability of harm).

  • Clinical Decision Support (CDS): Provide evidence-based recommendations at point-of-care.

• Outcome: Reduced clinician burden, improved documentation quality, better data for secondary analysis.

V. Patient Monitoring and Engagement

Remote Patient Monitoring (RPM) with AI

• System: Wearable sensors (ECG, SpO2, BP) → transmit data → Cloud/AI platform → Analyze → Alert clinician/patient.

• AI Role:

  1. Anomaly Detection: Identify vital sign deviations from personal baseline (using autoencoders, isolation forests).

  2. Trend Prediction: Forecast exacerbation (e.g., heart failure decompensation) using time-series models (LSTM, Prophet).

  3. False Alarm Reduction: Filter noise and non-critical events.

• Example: AI analyzing continuous ECG from a smartwatch to detect atrial fibrillation (AFib) episodes.

Wearable Health Technology

• Devices: Smartwatches, fitness trackers, patch sensors, smart rings.

• Sensors: PPG (photoplethysmography) for heart rate/BP, ECG electrodes, accelerometers, gyroscopes, temperature, glucose monitors.

• AI Connection: On-device or cloud-based ML models process raw sensor signals to derive clinical-grade metrics (e.g., sleep stages, arrhythmias, activity types). Enables continuous, real-world health assessment.

Virtual Health Assistants (VHAs) / Chatbots

• Function: Symptom checking, appointment scheduling, medication information, answering FAQs, mental health support.

• AI Tech: NLP (intent recognition, entity extraction), Dialogue Management, sometimes generative models (for conversational response).

• Benefit: 24/7 availability, reduces administrative load, improves patient access and engagement.

Medication Adherence

• Problem: ~50% of patients don't take meds as prescribed.

• AI Approaches:

  • Predictive: Identify patients at high risk of non-adherence using EHR data (models: logistic regression, XGBoost).

  • Intervention: Send personalized, timely reminders via app/SMS. Use reinforcement learning to optimize reminder timing/message.

  • Monitoring: Smart pill bottles (track openings), ingestible sensors (confirm ingestion).

Sentiment Analysis in Healthcare Context

• Goal: Extract subjective information (opinions, emotions, urgency) from patient-generated text (reviews, social media, support forum posts, patient-reported outcomes).

• Techniques: NLP - Lexicon-based (e.g., VADER) or ML/DL models (e.g., BERT fine-tuned on healthcare text).

• Applications:

  • Gauge patient satisfaction with care/services.

  • Identify unmet needs or common complaints.

  • Monitor mental health status from journal entries or social media (with ethical considerations).

  • Improve clinical documentation by analyzing physician notes for burnout indicators.

VI. Fundamentals of AI Model Development for Healthcare

Evaluation Metrics for Model Efficiency

• For Classification (e.g., disease detection):

  • Accuracy: $(TP+TN)/(Total)$. Misleading with imbalanced data.

  • Precision: $TP/(TP+FP)$. "When model says 'disease,' how often correct?" (Minimize false alarms).

  • Recall (Sensitivity): $TP/(TP+FN)$. "Of all actual disease cases, how many caught?" (Minimize missed cases).

  • F1-Score: Harmonic mean of Precision & Recall. $2*(Prec*Rec)/(Prec+Rec)$.

  • AUC-ROC: Area under ROC curve. Measures trade-off between TPR (Recall) and FPR across thresholds. \boxed{\text{Primary metric for binary classification in healthcare}}.

• For Risk Prediction/Prognosis (Time-to-event):

  • Concordance Index (C-index): Measures rank correlation between predicted risk and actual time-to-event. Analogue of AUC for survival models.

  • Calibration: Do predicted probabilities match observed frequencies? (Assessed with calibration plots, Brier score).

Techniques to Address Overfitting

• Definition: Model learns noise/training data specifics, performs poorly on new data.

• Techniques:

  1. Regularization: Add penalty to loss function.

     • L1 (Lasso): $$\displaystyle Loss + \lambda \sum |\beta_i| $$. Drives some coefficients to zero (feature selection).

     • L2 (Ridge): $$\displaystyle Loss + \lambda \sum \beta_i^2 $$. Shrinks coefficients.

  2. Cross-Validation (CV): k-fold CV is standard. Splits data into $k$ folds, trains on $k-1$, validates on 1, repeats. Provides robust performance estimate and helps tune hyperparameters.

  3. Pruning (for Trees): Remove branches with little predictive power (pre-pruning) or after full growth (post-pruning).

  4. Dropout (for Neural Networks): Randomly "drop" neurons during training to prevent co-adaptation.

  5. Early Stopping: Monitor validation loss during training; stop when it starts increasing.

  6. Ensemble Methods: Bagging (e.g., Random Forest) reduces variance; Boosting reduces bias.

  7. Increase Training Data: Often the most effective but least feasible.

[!TIP] In exams, always link the technique to the model type (e.g., "Dropout is for neural networks," "Pruning is for decision trees"). Cross-validation is universal for model assessment and selection.