AI in Healthcare and Medicine

AI in healthcare is the use of machine-learning systems, principally deep neural networks and, more recently, large language models, to support or automate tasks in clinical care, such as interpreting medical images, predicting patient deterioration, documenting consultations and answering clinical questions. The field spans products that hold regulatory authorisation and are in routine use, results from prospective clinical trials, and a larger body of retrospective or benchmark studies that have not been shown to change patient outcomes.

Medicine produces large volumes of structured and unstructured data (images, laboratory results, free-text notes) and many decisions that are repeated at scale, which makes it a natural target for statistical pattern recognition. It is also a domain where errors can harm patients, where deployment is gated by regulators, and where performance measured on a curated dataset often differs from performance in a particular hospital. For these reasons this article sorts evidence into three tiers and labels each claim accordingly:

A deployed product need not have strong outcome evidence, and a research demonstration can precede a product by years or never reach one. The three tiers should not be merged.

Rule-based clinical decision support dates to the expert-system era (see HIST-003 and HIST-004 for context on symbolic AI and its decline). The modern wave followed the demonstration that convolutional networks could match expert labelling on image tasks. A 2016 JAMA study reported a deep-learning algorithm for detecting diabetic retinopathy in retinal fundus photographs (Gulshan et al., 2016), and a 2017 Nature study reported classification of skin lesions on par with dermatologists in a test setting (Esteva et al., 2017). In 2018 an autonomous retinal screening device was tested in a pivotal trial (Abràmoff et al., 2018) and authorised by the US Food and Drug Administration (FDA) through the De Novo pathway on 2018-04-11 [VERIFICATION NEEDED: date taken from secondary press coverage; the FDA announcement itself was not retrieved]. Mammography followed with a 2020 international retrospective evaluation (McKinney et al., 2020) and a randomised screening trial reported in 2023 (Lång et al., 2023). From 2022 onward, large language models (LLMs) introduced a second line of work, from benchmark evaluations (Singhal et al., 2023) to documentation tools deployed in health systems (Tierney et al., 2024).

Discriminative models on images and signals. Most authorised devices are supervised classifiers or detectors trained on labelled examples. For a model $f_\theta$ and labelled data $(x_i, y_i)$, training minimises a loss such as cross-entropy,

Clinical performance is reported with sensitivity, specificity and the area under the ROC curve (AUC). Because screening prevalence is low, positive predictive value, which depends on prevalence by Bayes' rule, often matters more to a clinician than AUC.

Prediction models on electronic health records. Deterioration and sepsis models take structured features from records and output a risk score, which a hospital converts into an alert by choosing a threshold. The threshold, local prevalence and workflow determine usefulness as much as the model does.

Language models. LLMs are applied to question answering, summarisation and drafting. Ambient documentation tools combine speech recognition with an LLM that drafts a note for clinician review. Their failure modes (unsupported statements, omissions) are those described in CAP-LIMITS-001.

Why external validation matters. A model fitted at one site faces a shift in patient mix, devices and coding practice at another. Systematic reviews have found that most published deep-learning diagnostic studies are retrospective, few use external validation, and comparisons with clinicians are often poorly reported (Liu et al., 2019; Nagendran et al., 2020).

• Imaging AI (radiology, ophthalmology, dermatology, pathology): detection, triage and quantification. Radiology is the largest category of FDA-listed AI-enabled devices; a mid-2025 tally of the FDA list reported 956 of 1,247 devices (77%) in the radiology panel (FDA Law Blog, 2025). Counts change with each list update; see the Living Reference collections (FACT-001) for current figures.
• Autonomous versus assistive systems. An autonomous system issues a result without clinician interpretation of the image (the retinal screening device above). An assistive system presents a finding or prioritises a worklist for a clinician. Nearly all authorised imaging products are assistive.
• Clinical prediction on health records: early-warning scores, sepsis alerts, risk stratification for care management.
• Generative and conversational AI: documentation, patient messaging drafts, clinical question answering.
• Operational AI: scheduling, coding and prior-authorisation support, which do not make clinical decisions but affect access to care.

#6.1 Deployed products

• Autonomous diabetic-retinopathy screening. An FDA-authorised system analyses retinal photographs from a specified camera and returns a refer/re-screen result in primary-care settings without an eye specialist reading the image (Abràmoff et al., 2018).
• Regulatory footprint. The FDA publishes a list of AI-enabled devices authorised for marketing; it is dominated by imaging software, and an analysis of US and European approvals for 2015-20 documented the growth of such devices (Muehlematter et al., 2021). Authorisation attests to a regulatory standard of safety and effectiveness for the intended use, not to superiority over usual care.
• Ambient documentation. Health systems have rolled out AI scribes to draft visit notes. A published rollout at a large US health system reports clinician-experience outcomes (Tierney et al., 2024) [VERIFICATION NEEDED: study design and outcome measures were not read from the paper; treat as non-randomised until confirmed].

#6.2 Validated clinical results

• Pivotal trial, retinal screening. In a preregistered pivotal study enrolling 900 participants at primary-care sites, the system exceeded its prespecified endpoints, with sensitivity 87.2% (95% CI 81.8-91.2%), specificity 90.7% (95% CI 88.3-92.7%) and imageability 96.1% (Abràmoff et al., 2018). The comparator was reading-centre grading of wide-field photographs and optical coherence tomography. This is a diagnostic-accuracy result, not a patient-outcome result.
• Randomised mammography screening. The MASAI trial randomised 80,033 women to AI-supported screen reading or standard double reading. Its first, clinical-safety analysis (Lång et al., 2023) was a non-inferiority safety comparison; secondary summaries report higher cancer detection with AI support and a screen-reading workload reduction of roughly 44% [VERIFICATION NEEDED: exact detection rates and workload figures not read from the paper]. The primary endpoint, interval cancers, was not available at that analysis.

#6.3 Research demonstrations

• Dermatology. A deep network classified skin-lesion images with accuracy comparable to dermatologists on test images (Esteva et al., 2017). The evaluation was retrospective and image-based; it did not test use in clinical workflow.
• Mammography, retrospective. Using data from the UK and US, an AI system was compared with radiologists in a retrospective reader study (McKinney et al., 2020).
• Medical question answering. A prompted LLM was evaluated on multiple-choice and long-form medical questions, with clinician rating of answers, which exposed gaps relative to clinician answers (Singhal et al., 2023). Exam-style performance does not establish bedside safety.

Advantages. Consistent application of a rule at scale, availability where specialists are scarce (primary-care retinal screening), and possible workload reduction in high-volume screening.

Limitations and documented failures.

• Weak outcome evidence. Systematic reviews found that few deep-learning diagnostic studies were randomised or prospective, and that claims of clinician-level performance outran the study designs (Liu et al., 2019; Nagendran et al., 2020).
• External validation failure. An external validation of a widely deployed proprietary sepsis prediction model in 27,697 patients at one academic centre found poor discrimination (AUC 0.63) and a low share of sepsis cases identified, despite alerting on a substantial proportion of patients (Wong et al., 2021) [VERIFICATION NEEDED: AUC and sensitivity figures were checked against secondary summaries, not the full text].
• Algorithmic bias through proxy labels. A risk algorithm that used health-care cost as a proxy for health need under-identified Black patients with the same level of need; reformulating the target removed most of the bias (Obermeyer et al., 2019).
• Human-AI interaction. In a randomised trial of physicians solving diagnostic vignettes, access to an LLM did not significantly improve diagnostic reasoning compared with conventional resources, while the LLM alone scored higher than both physician groups (Goh et al., 2024). The result shows that model accuracy does not automatically transfer to a clinician-plus-tool team.

Unresolved questions. Whether AI screening reduces mortality or overdiagnosis, how performance drifts after deployment, how liability is allocated between clinician and vendor, and how to monitor generative tools whose output changes with each prompt.

As of 2026-10-15, the following can be stated at the stated evidence levels.

• Autonomous retinal screening: pivotal-trial accuracy evidence and regulatory authorisation exist (Well Supported for the stated task and camera).
• AI-supported mammography reading: randomised evidence on safety and screening metrics exists from one large trial programme; evidence on interval cancers and mortality was not confirmed in this draft (Preliminary).
• LLM clinical decision support: benchmark performance is high, randomised evidence of improved clinician decisions is mixed or absent (Contested).
• Equity and calibration problems are documented in deployed systems and require local monitoring (Well Supported).

Authoritative regulatory counts, cleared-device lists and guidance change frequently and are tracked in the Living Reference collections rather than here.

Near-term (0-3 years): continued growth of imaging and documentation products; more prospective trials and post-deployment monitoring requirements. Medium-term (3-10 years): evidence on outcomes (stage shift, mortality) from screening programmes; multimodal models integrating images, notes and laboratory data. Long-term (10+ years) scenarios are speculative and not assigned an evidence rating here.

• Foundational Concepts:
  • TECH-DL-001 Deep Learning: Training methods behind medical imaging models.
  • TECH-CNN-001 Convolutional Neural Networks: The architecture of most imaging classifiers.
  • TECH-LLM-001 Large Language Models: Basis of clinical question-answering and documentation tools.
• Related Methods & Technologies:
  • APP-SCI-001 AI in Scientific Discovery: Biomedical research applications, distinct from clinical care.
  • APP-PROT-001 Protein Structure Prediction (AlphaFold): Structural biology that feeds drug discovery.
  • CAP-LIMITS-001 Hallucination, Reliability and Calibration: Failure modes relevant to generative clinical tools.
• Advanced Implications & Future Trajectories:
  • GOV-LAW-001 AI Governance: Regulatory frameworks for medical AI.