28  Emerging AI Technologies in Healthcare

TipLearning Objectives

Foundation models, multimodal AI, and generative AI are transforming what’s possible. This chapter examines cutting-edge AI technologies and their potential clinical applications. You will learn to:

  • Understand foundation models and their healthcare applications
  • Evaluate multimodal AI systems integrating imaging, text, and genomics
  • Assess generative AI capabilities and limitations in clinical settings
  • Recognize edge AI, federated learning, and privacy-preserving technologies
  • Anticipate emerging trends: AI agents, real-time diagnostics, digital twins
  • Navigate the hype cycle—distinguish genuine advances from overpromises
  • Prepare for technologies likely to reshape clinical practice in 5-10 years

Essential for forward-thinking physicians and healthcare leaders.

Key Technologies on the Horizon:

Healthcare AI is transitioning from narrow, task-specific models to foundation models with broad capabilities. Key emerging technologies include:

  • Foundation models (e.g., Med-PaLM, GPT-4): Large language models trained on vast medical text, capable of answering clinical questions, generating notes, and assisting diagnosis
  • Multimodal AI: Systems integrating imaging, text, genomics, and wearable data for holistic assessment
  • Generative AI: Creating synthetic medical images for training, drafting clinical documentation, personalizing patient education
  • Edge AI: On-device processing for real-time diagnostics (e.g., point-of-care ultrasound interpretation)
  • Federated learning: Training AI across institutions without sharing patient data
  • Digital twins: Patient-specific computational models predicting disease trajectories and treatment responses

Reality Check: Many emerging technologies are in early stages. Pilots show promise, but prospective validation, regulatory approval, and real-world implementation lag behind hype. Physicians should engage with emerging technologies critically—neither dismissive nor uncritically enthusiastic.

The Path Forward: Medicine’s future involves AI partnerships, not replacement. Emerging technologies will augment clinical capabilities, but human judgment, ethical reasoning, and patient relationships remain irreplaceable.

28.1 Introduction

The past decade saw AI move from research labs to clinical deployment in narrow applications: detecting diabetic retinopathy, flagging intracranial hemorrhages, prioritizing radiology worklists. These successes established proof-of-concept—AI can match expert performance on well-defined tasks with abundant training data.

The next decade promises broader, more ambitious applications. Foundation models trained on billions of clinical datapoints may answer clinical questions, generate differential diagnoses, and assist documentation. Multimodal systems may integrate imaging, genomics, and real-time physiological data to predict patient trajectories. Generative AI may create personalized patient education, simulate clinical scenarios for training, or generate synthetic data to address dataset limitations.

This chapter examines emerging technologies reshaping healthcare AI—focusing on technologies likely to impact clinical practice within 5-10 years, while maintaining critical perspective on hype vs. reality.

28.2 Foundation Models in Healthcare

What are foundation models? Large neural networks trained on vast, diverse datasets, then fine-tuned for specific tasks. Examples: GPT-4 (OpenAI), PaLM 2 (Google), Claude (Anthropic). In healthcare, foundation models are trained on medical literature, clinical notes, guidelines, and patient data.

28.2.1 Med-PaLM and Medical Question-Answering

Med-PaLM (Google): Fine-tuned version of PaLM (Pathways Language Model) for medical applications (Singhal et al. 2023). Achieved passing scores on U.S. Medical Licensing Examination (USMLE) questions, approaching expert-level performance. Med-PaLM 2 improved further, demonstrating clinical reasoning capabilities.

Capabilities: - Answer medical questions with cited evidence - Generate differential diagnoses from clinical presentations - Explain complex medical concepts at appropriate literacy levels - Translate medical terminology for patients

Limitations: - Hallucinations: Confident but incorrect statements (fabricated citations, outdated information) - Lack of clinical context: May provide textbook answers without considering individual patient factors - Liability: Who is responsible when AI-generated advice is wrong? - Generalization: Training on Western medical literature may not reflect global diversity

Clinical use cases: - Clinical decision support (must be validated, not sole basis for decisions) - Patient education (requires physician review) - Medical training (simulated cases, exam preparation) - Literature review and synthesis (with human verification)

28.2.2 Generative AI for Clinical Documentation

Problem: Physicians spend 1-2 hours on documentation per hour of patient care. EHR burden contributes to burnout.

AI solutions: - Ambient clinical documentation: AI listens to patient encounters, generates draft notes (e.g., Nuance DAX, Abridge, Suki) - Structured data extraction: AI converts free-text notes to structured data for quality reporting, research - Auto-completion: Predictive text for common documentation tasks

Evidence: - Early pilots show time savings (30-50% documentation time reduction) - Physician satisfaction improves when AI drafts notes for review/editing - Accuracy concerns: hallucinations, missing key details, copy-forward errors

Critical considerations: - Review burden: Physicians must verify all AI-generated content—no time savings if review takes as long as writing - Legal implications: Signed note = physician attestation, even if AI-generated - Note bloat: AI may generate unnecessarily long, repetitive notes - Patient privacy: Conversations recorded and processed by third-party AI

28.3 Multimodal AI: Integrating Diverse Data Streams

Most current medical AI focuses on single modalities: images (radiology), text (NLP), or time-series (ECG). Real clinical reasoning integrates multiple data types: symptoms + imaging + labs + patient history + social determinants.

Multimodal AI systems combine imaging, text, genomics, wearables, and other data sources to generate holistic assessments.

28.3.1 Examples of Multimodal Medical AI

1. Vision-Language Models for Medical Imaging - Combine images with radiology reports, clinical notes - Example: CXR + presenting symptoms → differential diagnosis - Benefit: Contextualizes imaging findings (nodule in smoker vs. non-smoker)

2. Genomic-Clinical Integration - Combine genetic variants with EHR data (medications, diagnoses, labs) - Predict drug responses, disease risks, treatment outcomes - Example: Pharmacogenomic decision support integrating CYP2D6 genotype + current medications + liver function

3. Wearable Data + EHR Integration - Continuous physiological monitoring (heart rate, activity, sleep) + clinical data - Early warning systems for decompensation (heart failure, sepsis) - Example: Apple Watch AFib detection + EHR stroke risk factors → personalized anticoagulation recommendations

4. Pathology-Radiology-Genomics Integration - Oncology: combine histopathology (tumor morphology) + imaging (tumor burden) + genomics (mutations) → treatment recommendations - More accurate than single modality alone

28.3.2 Challenges for Multimodal AI

  • Data heterogeneity: Different formats, resolutions, timestamps
  • Missing data: Not all patients have all modalities (incomplete wearable data, missing labs)
  • Temporal alignment: Combining data from different timepoints (imaging from last week, labs from today)
  • Explainability: How to interpret complex, multi-source predictions?
  • Validation: Requires diverse datasets with all modalities—rare and expensive

28.4 Edge AI and Point-of-Care Diagnostics

Edge AI: Running AI models locally on devices (smartphones, ultrasound machines, wearables) rather than cloud servers.

Advantages: - Latency: Real-time processing without network delays - Privacy: Data stays on device, not transmitted to servers - Accessibility: Works in areas with limited internet connectivity

28.4.1 Clinical Applications

1. Point-of-Care Ultrasound (POCUS) AI - Smartphone-connected ultrasound devices (Butterfly iQ, Lumify) with AI guidance - AI assists probe positioning, image acquisition, interpretation - Use case: Emergency physicians with limited ultrasound training can assess cardiac function, detect pneumothorax, guide procedures

2. Smartphone-Based Diagnostics - Retinal imaging: Smartphone ophthalmoscope + AI → diabetic retinopathy screening in primary care - Dermatology: Smartphone photos + AI → melanoma detection - ECG: Smartphone-based ECG (AliveCor) + AI → AFib, QT prolongation, MI detection

3. Wearable AI - Continuous monitoring + on-device AI algorithms - Example: Smartwatch ECG + AI → AFib detection, fall detection, irregular rhythm alerts - Future: Predicting health events hours/days before symptoms (heart failure exacerbations, sepsis onset)

Challenges: - Limited computational resources: Complex models may be too large for devices - Battery constraints: Continuous AI processing drains batteries - Regulatory complexity: FDA considers device + AI together; updates require recertification - Accuracy trade-offs: On-device models often simplified vs. cloud-based versions

28.5 Federated Learning and Privacy-Preserving AI

Problem: Training accurate AI requires large, diverse datasets. But patient data cannot be freely shared across institutions due to privacy regulations (HIPAA, GDPR).

Solution: Federated learning—train AI models across multiple institutions without centralizing data.

28.5.1 How Federated Learning Works

  1. Central coordinating server distributes initial model to participating institutions
  2. Each institution trains model on local data (data never leaves institution)
  3. Only model updates (gradients, weights) sent back to central server
  4. Central server aggregates updates to improve global model
  5. Repeat until model converges

Advantages: - Patient privacy preserved (raw data never shared) - Models trained on larger, more diverse datasets than single institutions - Addresses dataset bias (includes data from diverse geographic, demographic populations)

Challenges: - Technical complexity: Requires standardized data formats, secure communication protocols - Computational cost: Each institution must run training - Governance: Who owns the model? How are contributions credited? - Security risks: Adversaries may infer sensitive information from model updates (differential privacy mitigates)

Real-world example: NVIDIA FLARE (Federated Learning Application Runtime Environment) used to train cancer detection models across 20+ institutions without sharing patient imaging data.

28.6 Generative AI: Synthesis, Simulation, and Creation

Generative AI creates new content—images, text, audio, video. In healthcare: synthetic medical images, personalized patient education, clinical scenario simulations.

28.6.1 Synthetic Medical Data

Problem: AI training requires massive labeled datasets. Medical data is scarce, expensive to label, privacy-restricted.

Solution: Generative AI creates synthetic medical images that resemble real images but don’t correspond to actual patients.

Methods: - Generative Adversarial Networks (GANs): Generator creates fake images, discriminator tries to distinguish real vs. fake—both improve iteratively - Diffusion models: Gradually denoise random noise into realistic images

Applications: - Data augmentation: Increase training dataset size - Privacy: Train models on synthetic data instead of real patient data - Rare diseases: Generate synthetic cases for conditions with limited real data

Limitations: - Synthetic data may not capture full complexity of real clinical data - Models trained solely on synthetic data may fail on real patients - Ethical concerns: creating “fake patients” for research?

28.6.2 AI-Generated Patient Education

LLMs can generate personalized explanations tailored to patient health literacy, language, cultural context.

Use case: Physician prescribes medication, AI generates handout explaining purpose, side effects, instructions at patient’s literacy level.

Challenges: - Accuracy: AI may provide incorrect medical information - Liability: Who is responsible if patient harmed by AI-generated advice? - Trust: Patients may prefer physician-written explanations

28.7 Digital Twins: Personalized Simulation Models

Digital twin: Virtual replica of a patient—computational model simulating physiology, disease progression, treatment responses.

Concept: Integrate patient’s imaging, labs, genomics, medical history into predictive model. Run simulations to forecast disease trajectories, test treatments virtually before real-world administration.

28.7.1 Potential Applications

1. Cardiovascular Disease - Patient-specific heart model from cardiac MRI - Simulate hemodynamics, predict heart failure progression - Test drug effects, device placement (pacemakers, stents) virtually

2. Oncology - Tumor model from imaging + genomics - Simulate growth patterns, predict treatment responses (chemo, radiation, immunotherapy) - Personalized treatment plans

3. Critical Care - Real-time physiological model from ICU monitors - Predict patient trajectory (septic shock, ARDS progression) - Optimize ventilator settings, fluid management

4. Surgical Planning - Patient-specific anatomical models for complex surgeries - Rehearse procedures, anticipate complications

28.7.2 Challenges for Digital Twins

  • Model complexity: Human physiology is extraordinarily complex—simplified models may not generalize
  • Data requirements: Need comprehensive, high-resolution patient data
  • Validation: How to validate predictions? Prospective trials costly and slow
  • Computational cost: Real-time simulations require significant processing power
  • Clinical integration: How to present predictions to clinicians? When to trust vs. override model?

Current status: Digital twins remain largely research-stage. Proof-of-concept studies show promise, but widespread clinical deployment years away.

28.8 AI Agents: Autonomous Clinical Decision-Making?

AI agents: Systems that perceive environment, make decisions, take actions autonomously (without human intervention at each step).

In healthcare: Could AI agents autonomously adjust medication doses, order labs, schedule appointments based on patient data?

28.8.1 Levels of Autonomy

  1. No autonomy: AI provides suggestions, physician decides and acts (current clinical decision support)
  2. Supervised autonomy: AI proposes actions, physician approves before execution
  3. Full autonomy: AI acts without human approval (rare in medicine, limited to low-risk tasks)

Example: Closed-loop insulin delivery - Continuous glucose monitor + insulin pump + AI algorithm - AI autonomously adjusts insulin delivery based on glucose levels - FDA-approved systems (e.g., Medtronic MiniMed 780G)—one of few examples of autonomous medical AI

Future possibilities: - Automated medication dose titration (anticoagulation, immunosuppression) - Lab ordering based on clinical algorithms (e.g., routine monitoring for chronic conditions) - Appointment scheduling based on symptom severity, test results

Barriers to AI agents in medicine: - Liability: Who is responsible when AI makes wrong decision? - Trust: Physicians and patients hesitant to cede control to algorithms - Complexity: Medicine involves nuanced judgment, weighing competing priorities—difficult to automate - Safety: Autonomous systems must be failsafe—no room for errors

Critical perspective: Full autonomy unlikely for high-stakes medical decisions. Supervised autonomy (AI proposes, human approves) more realistic for most clinical applications.

28.9 The Hype Cycle: Separating Signal from Noise

Emerging technologies attract hype—exaggerated claims, overoptimistic timelines, media sensationalism. Gartner’s Hype Cycle describes pattern: innovation trigger → peak of inflated expectations → trough of disillusionment → slope of enlightenment → plateau of productivity.

28.9.1 Evaluating Emerging Technologies

Ask critical questions:

  1. What problem does this solve? Technology without clear clinical need unlikely to succeed
  2. What’s the evidence? Retrospective studies? Prospective validation? Randomized trials? Real-world deployment data?
  3. Who benefits? Patients? Physicians? Hospitals? Payers? Vendors?
  4. What are failure modes? What happens when AI is wrong? Can errors be caught and corrected?
  5. What’s required for deployment? Regulatory approval? Workflow integration? Reimbursement? Training?
  6. What’s the timeline? Research prototype? FDA submission pending? Commercially available? Widely adopted?

Red flags for hype: - Claims of “revolutionary” technology without peer-reviewed evidence - Media coverage vastly exceeds published research - Vendor demonstrations on cherry-picked examples - No discussion of limitations, failure modes - Promises of imminent deployment without regulatory pathway

Green flags for genuine advances: - Peer-reviewed publications in high-quality journals - Prospective validation studies (not just retrospective) - Transparent reporting of limitations, failure cases - FDA clearance/approval (for devices) - Independent replication by multiple groups

28.10 Preparing for the AI-Augmented Future

What should physicians do now?

  1. Stay informed: Follow developments, read literature, attend conferences—but maintain critical perspective
  2. Engage with pilots: Participate in institutional AI implementations, provide feedback
  3. Demand evidence: Insist on prospective validation before widespread deployment
  4. Advocate for patients: Ensure emerging technologies serve patient welfare, not vendor profits
  5. Maintain core skills: AI will augment, not replace, clinical reasoning—foundational knowledge remains essential
  6. Shape policy: Physicians should lead discussions on AI governance, not defer to technologists alone

What emerging technologies are most likely to impact your practice in 5 years? - Foundation models for clinical decision support and documentation - Multimodal AI integrating imaging, labs, and clinical data - Edge AI for point-of-care diagnostics - Federated learning enabling multi-institutional AI training

What technologies remain further out (10+ years)? - Digital twins for personalized treatment simulation - Fully autonomous AI agents for clinical decision-making - General-purpose medical AI (single model handling all clinical tasks)

28.11 Conclusion

Emerging AI technologies promise to expand what’s clinically possible—answering complex questions, integrating vast datasets, personalizing treatments. But medicine has seen technological hype before (remember IBM Watson?). The gap between proof-of-concept and widespread clinical impact is vast.

Physicians should engage with emerging technologies—neither dismissive nor uncritically enthusiastic. Demand evidence. Insist on transparency. Center patient welfare. The future of medicine will be shaped by physicians who understand both the promise and limitations of AI.

Foundation models, multimodal AI, and digital twins represent genuine advances—but their path from research to clinical practice requires rigorous validation, thoughtful regulation, and continuous monitoring. Medicine’s future is neither fully human nor fully automated—it’s a partnership, guided by evidence and ethics.

28.12 References