How accurate are AI diagnostic models compared to clinicians?

Performance varies by modality and training data. In well-bounded tasks such as detecting diabetic retinopathy, classifying skin lesions, and triaging chest X-rays, top models have matched or exceeded specialist-level accuracy in controlled studies. However, real-world deployment introduces distribution shift, labelling inconsistency, and workflow friction that can widen the gap. AI performs best as a decision-support layer, not a standalone diagnostician.

What risks does AI introduce in clinical settings?

The primary concerns are automation bias, where clinicians over-rely on model outputs; dataset bias, where models trained on unrepresentative populations perform poorly for underrepresented groups; and the potential for adversarial manipulation of inputs. Regulatory frameworks like the EU AI Act classify most clinical AI as high-risk, requiring conformity assessments, post-market surveillance, and human oversight mechanisms. The FDA's regulatory approach continues to evolve as the technology matures.

Can AI reduce healthcare costs?

Evidence is mixed. AI-enabled early detection and triage optimisation can lower downstream treatment costs in specific applications. However, infrastructure, integration, and maintenance costs are substantial, and poorly implemented systems can increase clinician workload rather than reduce it. Cost-effectiveness depends heavily on the specific application and how well the tool fits existing clinical workflows. Rigorous health-economic evaluations remain limited for most deployed systems.

How is patient data protected when used to train health AI models?

Regulations including HIPAA, GDPR, and Malaysia's Personal Data Protection Act 2010 impose strict requirements on health data handling. Technical safeguards such as federated learning, where models train on data that never leaves the hospital, and differential privacy are emerging as standard approaches. However, research has demonstrated that trained neural networks can sometimes leak identifiable information about training examples through model inversion attacks. Data governance remains an active area of both research and regulation.

Home›Research›AI in Health

AI in Health Research

Diagnostic imaging, drug discovery, clinical decision support, and the regulatory landscape for AI in medicine

Overview

Artificial intelligence is no longer a speculative promise in medicine. It is a working tool in radiology departments, drug discovery labs, and primary care clinics around the world. The volume of published research has grown substantially since 2015, spanning diagnostic imaging, clinical decision support, robotic surgery, pharmacovigilance, and remote patient monitoring. What distinguishes the current wave from earlier cycles of enthusiasm is the breadth of real-world deployment: regulatory agencies on every continent have now approved AI-based medical devices, and several have reached routine clinical use.

The stakes are considerable. Healthcare systems face converging pressures from ageing populations, rising chronic disease burden, workforce shortages, and cost constraints. AI cannot solve these problems alone, but targeted applications are demonstrating measurable improvements in diagnostic speed, treatment precision, and operational efficiency. The key question has shifted from whether AI works in medicine to where it works best, under what conditions, and for whom.

This page presents verified research developments across the full spectrum of AI health applications. Each entry is sourced to peer-reviewed publications, regulatory filings, or institutional announcements. We distinguish between controlled study results and real-world deployment outcomes, and we note where evidence remains preliminary.

Key Developments

AlphaFold and Protein Structure Prediction

Google DeepMind's AlphaFold 2 achieved a median Global Distance Test score of 92.4 at the CASP14 competition in 2020, solving one of biology's longest-standing challenges: predicting protein three-dimensional structures from amino acid sequences. The AlphaFold Protein Structure Database now contains over 200 million predictions, accelerating drug target identification and enzyme engineering worldwide. Demis Hassabis and John Jumper were awarded the 2024 Nobel Prize in Chemistry alongside David Baker for their contributions to protein structure prediction. AlphaFold 3, announced in 2024 and published in Nature, expanded capabilities to predict protein-ligand, protein-DNA, and protein-RNA interactions.

FDA-Approved AI Medical Devices

The US Food and Drug Administration has authorised over 950 AI and machine-learning-enabled medical devices as of 2024. Approximately 77 percent of these devices are in the field of radiology, reflecting the maturity of image analysis applications. The first AI diagnostic device authorised by the FDA was IDx-DR in 2018, which detects diabetic retinopathy without requiring a specialist. The FDA published its AI/ML-Based Software as a Medical Device Action Plan in January 2021, outlining a regulatory framework for these technologies. Despite this progress, many AI tools used in healthcare globally remain under-regulated, with no standardised evaluation requirements in most jurisdictions.

AI-Assisted Radiology and Imaging

AI systems have demonstrated expert-level performance in detecting breast cancer from mammograms, with research published in Nature, Radiology, and The Lancet Digital Health. Google Health published a study in Nature in 2020 showing that its AI system outperformed radiologists in breast cancer screening on both UK and US datasets. Major medical imaging vendors including Siemens Healthineers, Philips, and GE Healthcare have integrated AI capabilities into their commercial imaging platforms. The UK National Health Service has piloted AI chest X-ray triage systems across multiple hospital trusts, aiming to reduce reporting turnaround times for critical findings.

Algorithmic Bias in Healthcare AI

A landmark study by Obermeyer and colleagues, published in Science in 2019, demonstrated that a widely used commercial algorithm systematically underestimated the illness severity of Black patients, effectively routing fewer Black patients to high-risk care management programmes despite having equal or greater medical need. The study catalysed increased scrutiny of algorithmic bias in medical AI across the industry. Subsequent research has documented bias in dermatology algorithms, which perform worse on darker skin tones, and in pulse oximetry models that can produce inaccurate readings for patients with darker skin pigmentation. The FDA has acknowledged the need for greater diversity in training datasets and has issued guidance on addressing bias in medical device development.

AI in Drug Discovery

Researchers at MIT used a graph neural network to screen over 107 million chemical compounds for antibiotic activity, identifying halicin, a structurally novel molecule effective against drug-resistant strains including Acinetobacter baumannii. Published in Cell in 2020, the work illustrated how AI can expand the chemical search space far beyond traditional high-throughput screening. Companies including Insilico Medicine, Recursion Pharmaceuticals, and BenevolentAI are actively applying AI to drug discovery pipelines. NVIDIA's BioNeMo platform provides generative AI tools for protein engineering and molecular design. Recursion Pharmaceuticals operates AI-powered robotic laboratories for high-throughput screening at scale.