A radiologist reviews dozens of mammograms a day, and the visual fatigue at the end of the shift degrades their accuracy. A deep learning algorithm does not tire, but neither does it understand the clinical context nor assume legal responsibility. AI-assisted medicine is not a replacement, but a collaboration between the two. This article describes how AI is applied to diagnostic imaging, genomics and drug discovery, and what the European healthcare regulatory framework — the Medical Devices Regulation and the AI Act — requires for a tool to reach the patient's bedside.
Diagnostic imaging: clinical computer vision
The most mature field is the analysis of medical images. Convolutional neural networks, and increasingly vision transformers, process X-rays, CT scans, MRIs and pathology slides for classification, detection and segmentation tasks. In screening for diabetic retinopathy, in detecting lung nodules on low-dose CT, or in delineating tumours to plan radiotherapy, these systems achieve a performance that, in controlled studies, matches that of specialists for specific tasks.
The technical key lies in how that performance is measured. Sensitivity — the proportion of patients correctly identified — and specificity — the proportion of healthy individuals correctly ruled out — trade off against each other, and the ROC curve with its area under the curve summarises the balance. In screening, very high sensitivity is preferred even at the cost of specificity, because missing a cancer is far more serious than generating an additional review. The decision threshold is therefore not a detail: it defines the clinical behaviour of the system.
Personalised medicine and genomics
Personalised medicine adapts prevention and treatment to each patient's molecular profile. Machine learning steps in by integrating heterogeneous data — genomic variants, gene expression, the electronic health record, biomarkers — to stratify risk and predict response to drugs. In oncology, models help predict which patients will respond to a specific immunotherapy based on the tumour's mutational load; in pharmacogenomics, they anticipate adverse reactions according to the patient's metaboliser genotype.
The dominant challenge is dimensionality: tens of thousands of genes against cohorts of a few hundred patients. Without rigorous regularisation and external validation on independent cohorts, the models overfit and discover spurious correlations that do not replicate. Reproducibility — not accuracy on a single set — is the criterion that separates a publishable finding from a clinically useful one.
The integration of multimodal data — imaging, genomics, the text of the clinical record and monitoring signals — is the most promising frontier and the most difficult. Each modality has its own scale, noise and missing values, and combining them requires architectures that learn to weight each source according to its reliability. When it works, the model captures nuances that no single source reveals: the same genetic marker can mean different things depending on the imaging and the patient's clinical course. When it fails, it is usually because of a naive integration that mixes data of very uneven quality and dilutes the useful signal in the noise.
Accelerated drug discovery
Designing a new drug takes more than a decade and billions of euros, with a very high failure rate. AI shortens several stages. In virtual screening, models predict which molecules from a vast library will bind to a therapeutic target, reducing laboratory assays. Protein structure prediction has transformed structural biology by inferring the three-dimensional fold from the amino acid sequence. And generative models propose new molecules with desired pharmacological properties, which are then synthesised and validated experimentally.
Caution is warranted: AI proposes candidates, but efficacy and safety are only demonstrated in clinical trials with people. No model replaces the regulatory phases.
An intermediate field growing rapidly is that of patient digital twins and in silico trials, where the response of a virtual organism to a compound is simulated before moving to the laboratory. These simulations do not replace the clinical trial, but they make it possible to discard early on candidates with a poor toxicity profile and to steer the design of studies toward the most promising hypotheses, which reduces the number of people exposed to molecules that will ultimately fail. The validity of these models depends entirely on the quality and representativeness of the biological data with which they are built.
Language models in clinical documentation
Beyond diagnosis, language models are making their way into a less glamorous but enormous task: the documentation burden. A doctor spends a substantial part of the working day writing reports, coding diagnoses and summarising medical records. Documentation assistants transcribe the conversation in the consultation, propose a draft of a structured report and suggest the corresponding codes, which the professional reviews and validates. The benefit is returning time to the relationship with the patient. The risk is hallucination: a model that invents a plausible but false clinical data point can have serious consequences, so these systems operate as support with mandatory human validation and never as an autonomous source of clinical information. The traceability of what the model generated versus what the clinician wrote is, here, a patient safety requirement.
| Application | Main technique | Maturity | Regulatory risk |
|---|---|---|---|
| Imaging screening | CNN / vision transformer | High (real clinical use) | Medical device |
| Risk stratification | Integrative models | Medium | High depending on use |
| Virtual drug screening | Graph-based learning | Medium-high (research) | Low (preclinical) |
| Clinical documentation assistants | Language models | Emerging | Variable, depending on the decision |
Regulatory framework: medical device and the AI Act
An algorithm intended to diagnose, prevent or treat a disease is, legally, a medical device. In the European Union it is governed by Regulation (EU) 2017/745 on Medical Devices, which requires CE marking, clinical evaluation and risk management throughout the entire life cycle. The ISO 13485 standard establishes the quality management system for medical device manufacturers, and ISO 14971 defines the risk management process. Software with a diagnostic function is usually classified in a medium or high risk class, which requires the involvement of a notified body.
To this is added the EU Artificial Intelligence Act, which classifies as high-risk the AI systems used as a medical device and imposes on them data governance, transparency, human oversight and activity logging. The processing of health data, considered a special category by the GDPR, requires reinforced safeguards: an explicit legal basis, an impact assessment and, where possible, anonymisation or pseudonymisation. The Spanish Agency for Medicines and Medical Devices supervises commercialisation in Spain.
Common mistakes and good practices in implementation
The first mistake is to train with data from a single hospital and deploy in another with different equipment, protocols and population: the model degrades through domain shift. The good practice is to validate externally on independent cohorts before any use. The second mistake is to ignore bias: if the training set under-represents a demographic group, the system will perform worse with it and amplify inequalities. The third is to present the model's output as a verdict rather than as decision support; qualified human oversight is a clinical and legal requirement, not a formality. Finally, performance in production must be monitored, because clinical protocols and imaging technology change over time.
Frequently asked questions
Can an algorithm issue a diagnosis without a doctor?
In European clinical practice, no. These tools are decision support and operate under the supervision of a healthcare professional, who retains responsibility. The AI Act reinforces this requirement for human oversight in high-risk systems.
What is the difference between sensitivity and specificity in this context?
Sensitivity measures how many patients the system correctly detects; specificity, how many healthy individuals it correctly rules out. In screening, sensitivity is prioritised so as not to miss cases, accepting more reviews of false positives.
How is patient data protected?
Health data is a special category under the GDPR and requires an explicit legal basis, an impact assessment and measures such as pseudonymisation. Access is restricted and logged, and models can be trained with techniques that minimise the exposure of identifiable data.
Does AI make healthcare more or less expensive?
It depends on the use case. In high-volume screening it can prioritise cases and reduce waiting times; in drug discovery it shortens preclinical stages. The real cost includes clinical validation, certification as a medical device and integration into hospital systems.
Conclusion
AI applied to healthcare advances faster in medical imaging and in the preclinical discovery of drugs than in autonomous diagnostic practice, and for good reasons: the regulatory bar is deliberately high when lives are at stake. The factor that decides whether a tool ends up being used is not the metric on a single test set, but reproducible external validation, the absence of demographic bias and the fit within the medical device framework and the AI Act. Personalised medicine will not arrive by replacing the clinician, but by giving them, at the moment of the decision, a second quantitative and traceable reading. At Summum we help healthcare centres and medical technology companies walk that path with methodological rigour and regulatory conformity.