AI in Healthcare and Medicine: A Foundation for Interdisciplinary Competence
Artificial intelligence is no longer a horizon technology in healthcare. It is actively reading ECGs, screening chest X-rays, flagging sepsis risk in ICUs, and transcribing clinical conversations into structured notes. Yet despite this diffusion into practice, most healthcare systems lack a shared understanding of what AI literacy should look like — and for whom.
This is a foundation article. Its purpose is to map the terrain: what AI is being used for in healthcare, what the evidence says, and — critically — which professional roles need to build competency in these areas, and why. This is not a call for every clinician to become a data scientist. It is a call for every health professional to become a fluent participant in a system increasingly shaped by machine intelligence.
The Current Landscape
The global AI in healthcare market was valued at approximately USD 20 billion in 2023 and is projected to exceed USD 200 billion by 2030, driven by improvements in compute, data availability, and model capability (Grand View Research, 2024). More significant than market size is the rate of clinical deployment — AI tools are embedded across radiology, pathology, cardiology, ICU monitoring, surgical assistance, and administrative workflows.
The evidence base has matured from proof-of-concept studies to large-scale prospective evaluations. A landmark 2022 review in Nature Medicine (Rajpurkar et al.) synthesised hundreds of clinical AI studies and identified consistent patterns: AI systems frequently reach or exceed specialist-level performance on narrow, well-defined tasks with high-quality imaging or structured data. The same review was careful to note that performance often degrades in real-world clinical environments, across patient populations underrepresented in training data, and in tasks requiring reasoning over incomplete information.
Understanding both what AI can and cannot do is the starting point for competence.
Key Application Domains
1. Medical Imaging and Computational Pathology
This is the most mature domain. Deep learning models trained on large annotated datasets now match or exceed expert radiologists at specific detection tasks: diabetic retinopathy (Gulshan et al., 2016, JAMA), skin cancer classification (Esteva et al., 2017, Nature), breast cancer screening (McKinney et al., 2020, Nature), and pneumonia detection from chest X-rays (Rajpurkar et al., 2017, arXiv).
In pathology, computational models analyse whole-slide images to assess tumour margins, grade malignancies, and predict molecular subtypes — tasks that previously required hours of pathologist review per case.
Clinical reality check: These models perform on curated test sets. Deployment across varied scanners, acquisition protocols, and patient demographics routinely reveals performance gaps. The UK NHS AI Lab’s prospective evaluations (Nair et al., 2023) and FDA’s AI/ML SaMD action plan both emphasise the need for post-market surveillance that most clinical AI deployments still lack.
2. Clinical Natural Language Processing and Documentation
Approximately 80% of clinical data exists as unstructured text — notes, discharge summaries, referral letters, operative reports. Clinical NLP extracts structured information from this text, and increasingly, LLMs generate it.
Ambient clinical intelligence tools — where an LLM listens to a consultation and produces a structured clinical note — are now commercially deployed (Nuance DAX, Nabla Copilot, Suki). Early evidence is promising: a 2023 evaluation at Stanford (Tierney et al., NEJM Catalyst) found physician documentation time fell significantly with ambient AI, though the concern remains about unreviewed AI-generated errors entering the medical record.
The broader NLP infrastructure — automated ICD coding, diagnostic entity extraction, medication reconciliation from clinical notes — is already running in the background of most large hospital information systems.
3. Clinical Decision Support and Predictive Analytics
Early warning scores for deterioration (sepsis, acute kidney injury, haemodynamic instability) represent one of the most widely deployed categories of clinical AI. The Epic Sepsis Model — integrated into the Epic EHR and used across hundreds of hospitals — became also one of the most scrutinised. A 2021 study by Wong et al. (JAMA Internal Medicine) found its real-world performance was substantially lower than reported in the original validation, with poor sensitivity and a high false positive rate generating “alert fatigue.”
This episode is instructive: it illustrates the gap between algorithmic performance on historical data and utility in a live clinical workflow. It also demonstrates that CDS tools are organisational interventions, not just software deployments — they change workflows, attention, and responsibility.
Positive evidence exists for well-scoped applications: AI-assisted antibiotic stewardship (Beaulieu-Jones et al., 2021), personalised chemotherapy dosing, and ICU discharge planning (Churpek et al., 2020). The pattern is consistent: tightly defined tasks, high-quality data, and close clinician-in-the-loop design outperform broad, “AI knows best” deployments.
4. Drug Discovery and Genomics
AlphaFold2’s near-complete solution of the protein structure prediction problem (Jumper et al., 2021, Nature, Science’s Breakthrough of the Year) reshaped computational biology. It demonstrated that deep learning could solve problems intractable by conventional approaches, and it has already accelerated drug target identification across diseases from malaria to Parkinson’s.
In genomics, AI models classify pathogenic variants, predict gene expression from DNA sequence, and integrate multi-omics data for stratified medicine. The integration of genomic AI into clinical practice — particularly in oncology and rare disease — is nascent but accelerating, with clinical genomics platforms now deploying ML pipelines directly in diagnostic workflows.
5. Remote Monitoring, Wearables, and Federated Health Data
Consumer wearables with medical-grade accuracy (Apple Watch ECG for atrial fibrillation detection, FDA-cleared continuous glucose monitors) place AI-driven diagnostics directly on patients’ wrists. The Apple Heart Study (Perez et al., 2019, NEJM) enrolled over 400,000 participants and demonstrated population-scale arrhythmia detection with a wearable device — a trial design that would have been impossible a decade ago.
Federated learning — training AI models across distributed hospital datasets without centralising patient data — is increasingly used to address data governance constraints. Projects including NVIDIA FLARE, the EU-funded MELLODDY project (drug discovery across pharmaceutical companies), and the NHS federated analytics platform are operational examples, not prototypes.
6. Administrative and Operational AI
The least glamorous but most immediately scalable domain. AI in prior authorisation, appointment scheduling, revenue cycle management, and patient communication reduces administrative burden estimated at 30–40% of healthcare expenditure in the US (Tseng et al., 2018, Annals of Internal Medicine). The gains here are real, measurable, and not subject to the safety stakes of clinical AI — which may be why adoption is faster.
Interdisciplinary Roles: Who Needs to Build Competency, and Why
AI in healthcare is not a single-discipline problem. It requires clinical knowledge, technical understanding, ethical judgement, organisational skill, and policy insight — distributed across professional roles that must work together. What follows is a role-specific mapping of required competencies, grounded in where each role intersects with AI systems.
Clinicians (Physicians, Nurses, Pharmacists)
Why AI competency matters: Clinicians are the last human checkpoint before AI outputs influence patient care. They must be able to critically evaluate AI-generated outputs, understand the conditions under which a system is reliable, and identify when algorithmic recommendations diverge from patient-specific context.
The EU AI Act (2024) classifies most clinical AI as “high-risk”, requiring human oversight. A physician who cannot meaningfully interpret an AI recommendation cannot provide that oversight. This is not analogous to understanding how an MRI scanner works — it is more like understanding when lab results should and should not be trusted.
Core competencies needed:
- AI output interpretation: sensitivity, specificity, positive predictive value in the deployment context
- Understanding training data limitations and distribution shift
- Recognising algorithmic bias as a clinical risk (e.g., pulse oximetry AI underperforming in darker skin tones; Sjoding et al., 2020, NEJM)
- Informed consent for AI-assisted care
- Documenting AI tool use in the medical record
Evidence: The Royal College of Physicians (UK) and the American Medical Association have both published frameworks for physician AI competency (2023). Medical licensing bodies in Germany (Bundesärztekammer) have begun incorporating digital health and AI literacy into continuing medical education requirements.
Health Informatics and Clinical Data Specialists
Why AI competency matters: This role sits closest to the technical architecture of deployed AI systems. Health informaticists design the data pipelines, interoperability frameworks (FHIR, openEHR), and governance structures that make AI possible — or prevent it from being deployed safely.
Core competencies needed:
- Data quality assessment and its implications for model performance
- AI/ML model evaluation methodology (train/test split, temporal validation, subgroup analysis)
- FHIR-based AI integration architectures
- Explainable AI (XAI) tools and their interpretation
- AI governance frameworks: FDA SaMD, EU AI Act, ISO 42001
Evidence: The AMIA (American Medical Informatics Association) 2024 competency framework for clinical informatics explicitly includes AI/ML skills as a core domain. IMIA and EFMI have issued joint recommendations calling for AI competency integration into health informatics curricula.
Healthcare Administrators and Hospital Executives
Why AI competency matters: AI procurement, implementation, and governance decisions are made at the organisational level. Administrators who lack AI literacy may approve systems that are technically impressive but clinically unsafe, or reject genuinely valuable tools due to misunderstood risk. They also bear responsibility for the ethical and legal framework within which AI operates.
Core competencies needed:
- Business case evaluation for AI: total cost of ownership, benefit measurement, failure mode planning
- Procurement criteria: algorithm documentation (model cards, datasheets for datasets), post-market surveillance requirements
- Staff change management for AI-augmented workflows
- Data governance, GDPR/HIPAA compliance for AI systems
- Risk stratification under AI Act and SaMD frameworks
Evidence: McKinsey (2023) found that AI implementation failure in healthcare is primarily organisational rather than technical — the limiting factors are workflow integration, staff adoption, and governance, not algorithm performance. The NHS AI Lab’s “AI in Health and Care Award” evaluation explicitly lists executive AI literacy as a key factor in successful implementation.
Biomedical Engineers and Medical Device Professionals
Why AI competency matters: Medical devices increasingly incorporate AI — from AI-enabled MRI reconstruction (reducing acquisition time by 50–70%) to surgical robots with computer vision assistance. Device engineers must understand both the engineering and regulatory dimensions of AI in safety-critical hardware.
Core competencies needed:
- AI/ML model validation for embedded medical systems
- Software as a Medical Device (SaMD) regulatory pathway
- Continuous learning systems and post-deployment monitoring
- AI safety standards: IEC 62304, ISO 14971 adapted for ML
- Adversarial robustness and failure mode engineering
Evidence: The FDA’s 2023 action plan for AI/ML-based SaMD and the International Medical Device Regulators Forum (IMDRF) guidance documents set explicit engineering requirements that necessitate embedded AI expertise in device development teams.
Public Health Officials and Epidemiologists
Why AI competency matters: AI is reshaping epidemiological surveillance — from infectious disease outbreak detection (BlueDot’s identification of COVID-19 spread patterns in January 2020) to chronic disease forecasting. Public health officials increasingly interact with AI-generated risk scores and population models in policy decisions.
Core competencies needed:
- Population-level AI model evaluation: equity, generalisability, ecological validity
- AI-driven surveillance systems and their limitations
- Synthetic data and privacy-preserving analytics for public health
- NLP for social media and unstructured public health signals
- AI bias and its implications for health equity
Evidence: WHO’s 2021 report Ethics and Governance of Artificial Intelligence for Health devotes specific attention to public health AI, warning that algorithmic systems risk amplifying existing health inequities if trained on historically biased data. The report calls for public health workforce competency as a global priority.
Medical Educators and Academic Researchers
Why AI competency matters: Medical educators shape the AI literacy of the next generation of clinicians and researchers. Academics set the evidentiary standards for clinical AI — they run the trials, evaluate the tools, and write the guidelines. Both roles require understanding not just how AI works but how to evaluate its evidence base critically.
Core competencies needed:
- Curriculum design for AI literacy (undergraduate through postgraduate)
- Critical appraisal of AI studies (TRIPOD+AI, SPIRIT-AI reporting standards)
- Research ethics for AI studies: consent, fairness, transparency
- Simulation-based AI training design
- Academic publishing standards for AI in medicine
Evidence: A 2021 study in JAMA Network Open (Parikh et al.) found that fewer than 10% of medical school curricula addressed AI at the time. Since then, institutions including Stanford, UCL, and Charité have integrated AI modules — but coverage remains inconsistent globally. TRIPOD+AI (Collins et al., 2024, BMJ) now provides the reporting standard for AI prediction model studies that researchers and reviewers must understand.
Bioethicists, Legal Professionals, and Policymakers
Why AI competency matters: The hardest AI problems in healthcare are not technical — they are normative. Who is liable when an AI-assisted decision leads to harm? How do we define meaningful informed consent for AI tools? How should algorithmic bias in diagnostic tools be regulated? These questions require people who understand both AI systems and legal/ethical frameworks.
Core competencies needed:
- Algorithmic accountability: auditability, explainability, right to explanation (GDPR Article 22)
- Liability frameworks for AI-assisted medical decisions
- Ethics of AI deployment in resource-limited settings
- AI Act compliance pathway and conformity assessment
- Participatory design and patient involvement in AI development
Evidence: The EU AI Act (Regulation 2024/1689) is the most comprehensive AI regulatory framework enacted to date. Its “high-risk AI system” classification covers most clinical AI tools and imposes technical, transparency, and human oversight requirements that cannot be met without competent professionals at every level of the healthcare system.
A Shared Competency Framework
Across all roles, the evidence points to a common set of foundational competencies that span profession-specific expertise:
| Competency Domain | Description |
|---|---|
| AI Literacy | Conceptual understanding of how AI systems learn, where they fail, and what their outputs mean |
| Data Literacy | Understanding of data quality, representation, bias, and governance |
| Critical Appraisal | Ability to evaluate AI study evidence using established reporting standards |
| Ethics and Law | Familiarity with regulatory frameworks, informed consent, and liability |
| Workflow Integration | Understanding how AI systems change human roles, responsibilities, and errors |
| Equity Awareness | Recognition of how AI systems can perpetuate or amplify health disparities |
No single professional requires deep expertise across all domains. But every professional involved in AI-enabled healthcare requires enough competency in each domain to participate meaningfully in its governance.
The Equity Problem Cannot Be an Afterthought
The most consistent finding in clinical AI research is that models perform unequally across patient populations. This is not a future risk — it is a documented, ongoing harm.
Obermeyer et al. (2019, Science) demonstrated that a widely-deployed commercial algorithm used to identify patients for care management programmes assigned substantially lower risk scores to Black patients than to equally sick White patients — an artefact of using healthcare costs as a proxy for health needs. The correction required identifying the bias, understanding the mechanism, and redesigning the system.
This required clinical knowledge (to identify the outcome misalignment), data science (to measure the bias), health systems knowledge (to understand why cost ≠ need), and ethical framing (to decide what to do about it). No single profession could have done it alone.
What Comes Next
The question is not whether AI will be embedded in healthcare. It already is. The question is whether the professionals working within that system are equipped to ensure it is deployed safely, equitably, and accountably.
The evidence is clear that AI produces real clinical benefit in specific, well-validated applications. The evidence is equally clear that AI produces real clinical harm when deployed without appropriate oversight, governance, and human expertise.
Building interdisciplinary AI competency is not a one-time training event. It is an ongoing organisational and educational commitment — one that must be embedded in professional curricula, continuing education frameworks, hospital governance structures, and research evaluation standards.
This article is a starting point. Future posts will go deeper into specific competency frameworks by role, tools for AI evaluation in clinical contexts, and what “trustworthy AI” looks like in practice at the point of care.
References:
- Collins GS, et al. (2024). TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ.
- Esteva A, et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature.
- Gulshan V, et al. (2016). Development and validation of a deep learning algorithm for detection of diabetic retinopathy. JAMA.
- Jumper J, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature.
- McKinney SM, et al. (2020). International evaluation of an AI system for breast cancer screening. Nature.
- Obermeyer Z, et al. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science.
- Perez MV, et al. (2019). Large-scale assessment of a smartwatch to identify atrial fibrillation. New England Journal of Medicine.
- Rajpurkar P, et al. (2022). AI in health and medicine. Nature Medicine.
- Sjoding MW, et al. (2020). Racial bias in pulse oximetry measurement. New England Journal of Medicine.
- WHO. (2021). Ethics and Governance of Artificial Intelligence for Health. World Health Organization.
- Wong A, et al. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine.