# Artificial Intelligence and Machine Learning in Preventive Medicine

## Overview

AI/ML applications in healthcare are growing rapidly, with >500 FDA-authorized AI/ML medical devices (majority in radiology and cardiology) In preventive medicine, AI/ML is applied to risk prediction, imaging interpretation, genomic analysis, disease surveillance, and population health management. Machine learning models can identify patterns in complex, high-dimensional data that traditional statistical methods may miss. Critical concerns include algorithmic bias, lack of transparency (black box models), data privacy, and regulatory oversight. Preventive medicine physicians must understand AI capabilities and limitations to evaluate, implement, and govern these tools.

## Fundamentals of AI/ML in Healthcare

### Key Concepts

**Artificial Intelligence (AI)**: broad field of computer science focused on creating systems that perform tasks requiring human intelligence. **Machine Learning (ML)**: subset of AI where algorithms learn from data to make predictions without explicit programming. **Deep Learning (DL)**: subset of ML using neural networks with multiple layers; excels at image/signal/text processing. **Natural Language Processing (NLP)**: AI techniques for understanding and generating human language; used for clinical note analysis, chatbots. **Training data**: historical data used to build models; quality and representativeness are critical. **Validation**: testing model performance on independent data not used during training. **External validation**: testing on data from different settings/populations; essential before clinical deployment.

### Types of ML Approaches

**Supervised learning**: learns from labeled data (e.g., predicting diabetes from clinical variables); classification (disease/no disease) or regression (continuous outcome) **Unsupervised learning**: identifies patterns in unlabeled data (e.g., patient clustering, phenotype discovery) **Reinforcement learning**: learns through interaction with environment and reward signals (e.g., optimizing treatment protocols) **Transfer learning**: adapts a model trained on one task to a different but related task (e.g., retraining imaging models for new pathology)

| ML Approach | Description | Example Application |
|---|---|---|
| Supervised learning | Learns from labeled data | Predicting diabetes from clinical variables |
| Unsupervised learning | Identifies patterns in unlabeled data | Patient clustering, phenotype discovery |
| Reinforcement learning | Learns through interaction and reward signals | Optimizing treatment protocols |
| Deep learning | Neural networks with multiple layers | Medical image interpretation |
| Natural language processing | Understanding/generating human language | Clinical note analysis, SDOH extraction |
| Transfer learning | Adapts model from one task to a related task | Retraining imaging models for new pathology |

## Applications in Preventive Medicine

### Risk Prediction Models

Predicting diabetes onset, cardiovascular events, hospital readmissions, suicide risk, sepsis. ML models (random forests, gradient boosting, neural networks) may outperform traditional logistic regression for complex, non-linear relationships. Example: ML-based CVD risk calculators incorporating EHR variables beyond traditional Framingham/PCE inputs. Concerns: marginal improvement over simpler models in many cases; complexity without proportional clinical benefit.

### Medical Imaging

**Radiology**: lung nodule detection on LDCT, mammography interpretation (AI-aided readers), diabetic retinopathy screening. **Dermatology**: melanoma vs. benign lesion classification from photographs. **Pathology**: digital pathology for cancer grading and biomarker detection. FDA-cleared examples: IDx-DR (autonomous diabetic retinopathy screening), AI-aided mammography readers (Lunit, iCAD, Hologic) Potential to extend screening to underserved areas where specialist access is limited.

### Genomics and Precision Prevention

Polygenic risk scores (PRS): AI-assisted computation of genetic risk from millions of variants. Pharmacogenomics: predicting drug metabolism and response. Cancer risk: integrating genomic, clinical, and lifestyle data for personalized screening intervals. Limitation: most genomic ML models trained primarily on European-descent populations; poor generalizability to diverse populations.

### Disease Surveillance and Outbreak Detection

Syndromic surveillance: AI algorithms detect anomalous patterns in ED data, lab results, or social media. Nowcasting: real-time estimation of disease incidence (e.g., Google Flu Trends -- illustrative failure; more sophisticated successors) Wastewater-based epidemiology: ML analysis of pathogen genomic sequences in sewage for community-level surveillance. COVID-19: AI used for case forecasting, contact tracing optimization, and variant tracking.

### Population Health Management

Identifying high-risk patients for proactive intervention (care management, transitional care) Predicting no-shows and optimizing appointment scheduling. NLP for extracting social determinants from clinical notes. Automating quality measure calculation from EHR data.

## Algorithmic Bias and Health Equity

### Sources of Bias

**Training data bias**: if historical data reflects existing disparities (e.g., differential referral patterns by race), the model learns and perpetuates those disparities. **Label bias**: biased outcome labels (e.g., using healthcare costs as proxy for healthcare needs disadvantages populations with lower healthcare utilization) **Measurement bias**: differential data quality across populations (missing data, coding differences) **Representation bias**: underrepresentation of minority populations in training datasets. **Automation bias**: clinicians may over-rely on AI recommendations without critical evaluation.

### Case Study: Optum/UnitedHealth Algorithm

Widely used algorithm predicted healthcare needs using healthcare costs as the outcome variable. Black patients had lower costs than White patients at the same level of illness (due to access barriers and historical underinvestment) Result: algorithm systematically underestimated illness severity in Black patients, reducing their access to care management programs. Correcting for this bias would have more than doubled the proportion of Black patients flagged for intervention. Published in Science (Obermeyer, 2019); landmark demonstration of algorithmic racism.

### Mitigation Strategies

Diverse, representative training data. Bias auditing: testing model performance across demographic subgroups (fairness metrics) Choosing appropriate outcome variables (avoid proxies that encode existing disparities) Multidisciplinary oversight: including ethicists, community representatives, and equity experts in AI governance. Transparency: publishing model performance disaggregated by race, sex, age, SES.

## Regulatory Framework

### FDA Oversight

FDA regulates AI/ML as Software as a Medical Device (SaMD) >500 authorized AI/ML devices; most through 510(k) pathway (substantial equivalence) De Novo pathway: for novel AI devices without predicates. FDA action plan (2021): proposed framework for continuously learning AI (models that update over time) Predetermined Change Control Plan: allows manufacturers to describe anticipated modifications without new FDA submissions.

### Current Gaps

No requirement for external validation in diverse populations before clearance. Limited post-market surveillance of AI performance degradation (model drift) Variable transparency: many FDA-cleared AI tools are black boxes with undisclosed training data composition. No standardized requirements for bias testing or health equity impact assessment.

<image>A taxonomy diagram showing the AI/ML landscape in healthcare: the outer ring shows AI applications by domain (radiology, pathology, genomics, surveillance, population health, clinical decision support). The middle ring shows ML technique types (supervised, unsupervised, deep learning, NLP, reinforcement learning). The inner core shows foundational requirements: training data, validation, external validation, bias auditing, and regulatory clearance. AI in healthcare education illustration.</image>

<image>A case study infographic of the Optum/UnitedHealth algorithm bias: two patient silhouettes (Black and White) with the same illness severity but different healthcare costs (due to access disparities). The algorithm, trained on cost data, scores the Black patient as lower risk and the White patient as higher risk, resulting in differential care management enrollment. A corrected model using health-based rather than cost-based outcomes produces equitable risk scores. Algorithmic bias education illustration.</image>

<image>A flowchart showing the FDA regulatory pathway for AI/ML medical devices: starting with classification (Class I, II, III), then pathway selection (510(k), De Novo, PMA), followed by performance evaluation requirements (analytical validation, clinical validation). A branch shows the predetermined change control plan for adaptive/learning algorithms. Key regulatory gaps (bias testing, post-market surveillance, transparency requirements) are highlighted in callout boxes. AI regulation education illustration.</image>

## Clinical Pearls

The Optum algorithm bias case demonstrates that using healthcare cost as a proxy for health need systematically disadvantages populations facing access barriers -- preventive medicine physicians must critically evaluate what outcome an AI model is actually optimizing. AI-aided mammography and diabetic retinopathy screening have the potential to extend specialist-level screening to underserved settings -- this is where the equity promise of AI is strongest. Polygenic risk scores are currently limited by their derivation from predominantly European-descent cohorts -- applying them to diverse populations without recalibration may widen rather than narrow disparities. For boards: understand the difference between supervised/unsupervised learning, the concept of external validation, sources of algorithmic bias, and the FDA SaMD regulatory framework. Model drift (degradation of performance over time as patient populations and clinical practices change) is a real-world concern that requires ongoing monitoring after AI deployment.

## References

- Obermeyer Z, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453.
- Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56.
- FDA. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device Action Plan. FDA; 2021.
- Rajkomar A, et al. Machine learning in medicine. N Engl J Med. 2019;380(14):1347-1358.
- Vyas DA, et al. Hidden in plain sight -- reconsidering the use of race correction in clinical algorithms. N Engl J Med. 2020;383(9):874-882.