AI in Clinical Medicine

Artificial intelligence (AI) is a field of computer science that uses algorithms to replicate human intelligence tasks such as pattern recognition, decision-making and problem-solving through complex datasets. ^[1] In clinical care, AI technologies including machine learning (such as neural networks and decision trees), deep learning and natural language processing are being integrated into clinical decision support systems for diagnostics, treatment recommendations, risk prediction and documentation. ^[2] Large language models (LLMs) are AI tools based on transformer architectures that can respond to free-text queries without being specifically trained on the task in question. ^{[3] [4]} The field spans disease detection, personalized care, drug discovery, predictive analytics, telemedicine and wearable health technologies, drawing on electronic health records, medical imaging and genomic profiles. ^[5] Multimodal AI integrates diverse data such as images and genetic data, and AI-driven tools are expected to support proactive, personalized healthcare through continuous monitoring and multiscale forecasting. ^[6] Medical digital twins, constantly updated virtual patient models that combine multimodal data, AI and mechanistic modelling, are proposed as a way to enable precision medicine, although there is not yet a consensus on their fundamental make-up. ^[7]

In medical imaging, radiomics builds machine-learning models from handcrafted features derived from mathematical formulas, whereas deep learning convolutional neural networks extract features automatically from the images. ^[8] Deep learning convolutional neural networks applied to single, continuous and intermittent ECG leads can identify subtle patterns not readily apparent to human interpreters, which is why AI-enhanced ECG algorithms are described as non-invasive biomarkers of cardiovascular disease. ^[9] For breast cancer screening on digital mammography and digital breast tomosynthesis, deep learning systems are a distinct technology from the earlier computer-aided detection (CAD) systems and require independent validation. ^[10] Generative LLM chatbots are produced by fine-tuning a large language model, which lets them generate text in response to clinical, educational and research queries. ^[4] LLMs can process unstructured text, including patient data and medical guidelines, and integrate several information sources, with accuracy reported to be highest when they are augmented with retrieval-augmented generation. ^[11]

In a systematic review and meta-analysis of 39 studies of hip fracture radiographs and postoperative outcomes, the odds ratio for diagnostic error of machine-learning models compared with clinicians was 0.79 (95% CI, 0.48-1.31), a performance the authors judged comparable with expert radiologists and surgeons. ^[13] In the same meta-analysis, machine-learning models for hip fracture diagnosis had a mean sensitivity of 89.3% and specificity of 87.5%, but for predicting mortality after surgery their mean area under the curve of 0.84 versus 0.79 for traditional regression models did not provide substantial benefit. ^[13] A systematic review of AI algorithms for early skin cancer detection found reasonable mean diagnostic accuracy for melanoma, squamous cell carcinoma and basal cell carcinoma, but because few studies trained or tested algorithms in low-prevalence populations, adoption in community and primary care could not be recommended. ^[14] In a single-blind randomized clinical trial in which physicians trained in family, internal or emergency medicine worked through clinical vignettes, access to an LLM did not significantly improve diagnostic reasoning compared with conventional resources (median score per case 76% versus 74%, adjusted difference 2 percentage points). ^[15] In that trial's exploratory analysis, the LLM used alone scored 16 percentage points higher than the conventional-resources group (95% CI, 2-30), which the authors interpret as a need for workforce development to realize physician-AI collaboration. ^[15]

Real-world validation remains scarce: most AI studies in clinical infection practice lack real-world validation or clinical utility metrics, and in a systematic review of AI in nursing and midwifery only 8 (7.14%) studies reported actual benefits in real-world settings. ^{[17] [18]} Of 519 studies evaluating LLMs in health care, only 5% used real patient care data, and fairness, bias and toxicity were assessed in 15.8% while accuracy was the primary dimension in 95.4%. ^[19] Unbalanced training datasets produce biases based on age, sex and race, as shown for AI-enhanced ECG, where model performance falls when diverse demographics are inadequately represented. ^[9] Proposed fairness safeguards include diverse and representative data, algorithm audits and ongoing monitoring, alongside clear responsibility frameworks for manufacturers, healthcare institutions and professionals. ^{[20] [21]} Generalizability is limited when training data come from specialized centers or from in vitro and ex vivo datasets that do not replicate clinical environments. ^[22]