Reading time: 32 minutes
Artificial intelligence
and tomorrow's medicine
Clinical, organizational, economic, ethical and professional impacts — a full prospective analysis
Artificial intelligence is not an additional medical innovation comparable to the arrival of a new imaging device or a new drug. It acts as a transversal technology, capable of simultaneously modifying disease detection, the interpretation of examinations, the formulation of differential diagnoses, the customisation of treatments, the continuous monitoring of patients, pharmaceutical research, the organisation of hospitals, medical documentation, the training of professionals and the relationship between the patient, the doctor and the care system.
- Disease detection
- Interpretation of examinations
- The formulation of differential diagnoses
- Customization of processing
- Continuous monitoring of patients
- Pharmaceutical research
- Organisation of hospitals
- Medical documentation
- Training of professionals
- The relationship between the patient, the doctor and the care system
Radiology is one of the first highly exposed specialties, as it produces considerable volumes of standardized numerical data. — X-rays, scanners, MRIs and mammograms — particularly suitable for machine learning. But the impact goes far beyond imaging: pathology, dermatology, ophthalmology, cardiology, oncology, genomics, surgery, general medicine, psychiatry and clinical research are also involved.
IA is likely to replace doctors less than it will replace certain tasks now performed by doctors. On the other hand, professionals who do not use AI may be gradually marginalized by those who can integrate it into rigorous clinical practice.
Summary of Article
Understanding the artificial intelligence revolution in Medicine
Chapter I What artificial intelligence are we talking about in medicine?
The term « artificial intelligence » covers several technologies with different maturity levels, objectives and risks.
| Technology | Main function | Medical examples |
|---|---|---|
| Automatic learning | Identifying statistical relationships | Prediction of the risk of re-hospitalisation |
| Deep learning | Recognizing complex structures | Detection of a tumour on a mammogram |
| Artificial vision | Analyze images or videos | Radiology, dermatology, endoscopy |
| Treatment of natural language | Understanding and structuring texts | Analysis of medical records and records |
| IA generative | Produce text, images or syntheses | Writing letters and clinical notes |
| Multimodal models | Combine text, image, biology and clinical | Integrated diagnostic in oncology |
| IA predictive | Estimating the likelihood of an event | Risk of infarction, sepsis or complication |
| Prescriptive IA | Propose therapeutic behaviour | Treatment choice or dosage adjustment |
| Smart robotics | Assist or automate a gesture | Surgery, rehabilitation, hospital logistics |
| Self-employed | Perform a sequence of tasks | Coordination of examinations, preparation of a dossier |
Large multimodal models are particularly strategic: they can jointly process language, imaging, biological data and other clinical information. Their power, however, increases the stakes of governance, deexplicability, security and validation.
Chapter II Why is radiology in the forefront?
1. A fully digitized specialty
- Data already digital
- Relatively standardized images
- Very large volumes of examinations
- Anomalies often visually detectable
- Formalized diagnostic classifications
- Numerous previous reviews available
- High interpretation delay issue
2. What AI already knows how to do
- Detect pulmonary nodules, embolis, fractures or bleeding
- Analyze mammograms
- Segmenting and measuring a tumor
- Automatically compare multiple exams
- Prioritizing urgent reviews
- Produce an initial version of the report
- Quantify progression of pathology
- Control the technical quality of an image
3. IA as second reader
- Analysis of the examination.
- Report suspicious areas.
- Possible allocation of a probability score.
- Provision of quantitative measures.
- Priority of urgent cases for the radiologist.
4. Will the radiologist disappear?
This assumption appears unlikely in the medium term. Radiological interpretation is not limited to recognizing a form: it requires verification of the quality of the examination, the integration of the history, the temporal comparison, the assessment of the consistency between the image and the symptoms, the management of the fortuitous discoveries, the dialogue with clinicians, the carrying out of interventional gestures and the assumption of medical responsibility.
The profession will evolve towards more validation, complex case resolution, multimodal integration, intervention, tool control and management of differences between AI and human appreciation.
Chapter III Other medical specialties already transformed
The impact of artificial intelligence is not limited to imaging. Eight other clinical fields are already undergoing significant transformations.
1. Anatomopathology
The scanning of blades allows AI to identify tumour areas, count cells, characterize the aggressiveness of a tumour, evaluate certain biomarkers, quantify protein expression and detect microscopic metastases. The pathologist remains responsible for the final interpretation.
2. Dermatology
The AI can classify lesions, estimate a risk of melanoma, monitor the evolution of a nævus, or direct an urgent consultation. The biases associated with skin colour, lighting and under-representation of certain populations are major.
3. Ophthalmology
Automated background analysis facilitates the detection of diabetic retinopathy, glaucoma and macular degeneration. Some systems may recommend referral without immediate intervention by an ophthalmologist under strict protocol.
4. Cardiology
Applications cover ECG analysis, detection of rhythm disorders, risk assessment of heart failure, ultrasound analysis and connected object monitoring.
5. Oncology
The AIA combines imaging, anatomopathology, genomics, biology and therapeutic data to detect, characterize and monitor cancers, predict recurrence or toxicity, select treatments and optimize radiation therapy.
6. Neurology
Applications include stroke, brain MRI, epilepsy, neurodegenerative diseases and analysis of speech, walking, striking or sleep as digital biomarkers.
7. Psychiatry and mental health
The AIA can support screening and follow-up, but presents specific risks: misinterpretation, excessive surveillance, damage to privacy, confusion between conversational support and psychotherapy, inadequate emergency management.
8. General medicine
General medicine will be profoundly transformed through summary of the file, preparation of the consultation, differential diagnoses, drug verification, writing and follow-up. However, it remains based on complex, multi-morbid and social situations that are difficult to reduce to structured data.
Chapter IV Generative LA: the silent revolution of medical work
1. Automated documentation
Ambient AIA systems can listen to a consultation, with the patient's agreement, and then produce a structured note, a clinical summary, a list of problems, a mail, follow-up instructions or a proposed prescription. This technology can reduce the documentary burden but requires systematic human validation.
2. Automated summary of the file
- Reconstitute medical chronology
- Extract relevant background
- Identify abnormal results
- Remove medicines and allergies
- Detect missing exams
- Summarize hospitalizations
- Identify inconsistencies
3. Diagnostic aid
Generative LAA may offer differential diagnoses, but a generalistic model is not a validated medical device. It can produce a convincing but erroneous answer, invent a reference, neglect a rare diagnosis, overestimate a hypothesis or vary depending on the formulation of the question.
Towards predictive medicine, custom and increased
Chapter V From curative medicine to predictive medicine
1. Risk prediction
Based on records, analyses, images, genetic data and connected objects, AI can estimate the risks of diabetes, infarction, stroke, renal failure, cardiac decompensation, infection, postoperative complication, re-hospitalization, cancer or fall.
2. Detection before symptoms
Algorithms can detect weak signals — subtle anomalies on an image, biological variations, change of voice, disturbance of walking or sleep — prior to symptoms. The symmetrical risk is overdiagnosis.
3. Digital twins
A digital twin is a dynamic computer model representing certain anatomical, physiological, genomic and therapeutic aspects of a patient. It could simulate dosage, surgery, radiation therapy or a change of strategy before the actual intervention.
Chapter VI A personalized medicine on a very large scale
Evidence-based medicine often produces average results. The purpose of the AIA is to identify the most relevant treatment for a particular patient, its optimal dose, its risk of toxicity and when a strategy should be changed.
The decision can combine imaging, genomics, proteomics, microbiota, biology, clinical history, behavior, environment, socio-economic data and patient preferences. This ambition requires reliable, representative, interoperable and legally accessible data.
Chapter VII AIA in surgery and medical gestures
- Pre-operative planning
- Anatomical reconstruction 3D
- Identification of risk structures
- Guidance of the gesture
- Movement stabilization
- Video analysis
- Detection of bleeding
- Quality assessment
- Prediction of complications
- Simulation training
The most likely evolution is gradual: visual assistance, trajectory recommendation, sub-task automation, partial autonomy under supervision. Fully autonomous surgery remains limited by anatomical variability, unforeseen events, responsibility and acceptability.
Chapter VIII Transformation of medical and pharmaceutical research
1. Discovery of medicines
AIA can accelerate target identification, protein modelling, molecular selection, toxicity prediction, repositioning of drugs, and design of new molecules. It does not remove biological testing but reduces the number of unnecessary candidates.
2. Clinical trials
- Identification of eligible patients
- Selection of centres
- Analysis of inclusion criteria
- Adverse event detection
- Quality monitoring
- Establishment of external control groups
The methodological difficulty is to evaluate an evolutionary algorithm. Devices require dynamic testing and post-deployment monitoring.
Patients, health systems and clinical risks
Chapter IX The patient of tomorrow: better informed, but also more exposed
1. Prediagnosis by the patient
Patients already use conversational assistants to understand a symptom, prepare a consultation, interpret a result, ask for a second opinion or redraft a report. Consultation will become more of an arbitration and validation space.
2. The risk of a false sense of competence
- False reinsurance
- Delay in consultation
- Self-medication
- Misinterpretation
- Anxiety
- Unnecessary reviews
- Exposure of personal data
3. Towards renewed consent
The patient should be informed when AI intervenes substantially in the analysis of an examination, a therapeutic recommendation, the recording of a consultation, the prediction of a risk or the prioritization of care.
Chapter X Expected benefits for health systems
1. Productivity gains
AIA can reduce time spent writing, coding, screening, seeking information and planning. However, a technical gain can be absorbed by verification, false positives, licenses, integration and maintenance.
2. Reduction of medical deserts
- Teleexpertise
- Emergency sorting
- Initial interpretation
- Screening protocols
- Assistance to nurses
- Continuing training
- Medical translation
- Remote tracking
The AIA improves the use of scarce resources but does not replace infrastructure, medicines, professionals, or healthcare systems.
3. Improving continuity
Systems can detect treatment failure, missed appointment, unconsulted outcome, interaction, lack of follow-up or risk of re-hospitalization.
Chapter XI Major clinical risks
Seven categories of clinical risks accompany the integration of artificial intelligence into medical practice.
1. Bias
An AI learns from the available data. Under-representation of certain populations, baseline errors, coding practices or differences in access to care can produce inequitable performance.
2. Change of context
An algorithm validated in one centre can lose performance in another facility, on another device or when prevalence changes: it is the distribution drift.
3. False positives and false negatives
The consequences can be unnecessary testing, biopsies, anxiety, therapeutic delay or false reinsurance.
4. Automation Bias
The professional may give excessive weight to the algorithmic recommendation or, conversely, ignore it by habit.
5. Erosion of skills
Excessive dependence can reduce the ability to interpret, synthesize or detect an error without assistance.
6. Opacity
A medical decision must remain explicable: data taken into account, limitations, uncertainties and clinical consistency.
7. Cybersecurity
Theft, ransomware, falsification of results, model alteration or intrusion into a connected device become direct clinical risks.
Chapter XII Health data, sovereignty and governance
The strategic value lies both in the data and in the algorithm: volume, quality, representativeness, longitudinality and interoperability are decisive.
1. The data paradox
The more centralized the data are to improve models, the greater the risk of re-identification, leakage, discrimination, unplanned secondary use and commercial dependence.
2. European Health Data Area
The European framework, which entered into force in March 2025, aims to facilitate citizens' access to their data and their supervised reuse for research, innovation and public policies.
3. Risk of technological dependence
- Ownership of derived data
- Model auditability
- Transferability
- Continuity in the event of supplier failure
- Re-use for training
- Contract lock
Chapter XIII Legal responsibility: who is responsible for the error?
In the event of error, the liability may relate to the doctor, institution, manufacturer, integrator, data provider or publisher. The traditional model, in which the professional retains the final decision, becomes delicate when the tool is opaque, imposed or evolving.
1. Does the doctor remain responsible?
The doctor cannot reasonably be obliged to use a system and held solely responsible for a function which he cannot understand or modify. Governance must clearly allocate responsibilities.
2. The European framework
Systems embedded in certain medical devices can be classified as high risk and subject to risk management requirements, data quality, documentation, traceability, human supervision, robustness, cybersecurity and post-market monitoring.
3. Medical device and generalist chatbot
A tool designed, tested and certified for a medical function should not be confused with a generalistic model capable of discussing health without specific clinical validation.
Chapter XIV Will the medical professions be destroyed?
Most automated tasks
- Repeated and standardized tasks
- Recognition of grounds
- Sorting folders
- Transcript
- Executive summary
- Codification
- Calculation of scores
- Control of requirements
Less automated tasks
- Empathy
- Physical examination
- Contextual judgment
- Ethical arbitration
- Negotiation
- Uncertainty Management
- Human coordination
- Complex gesture
- Communication of bad news
Likely impact matrix
| Occupation | Task exposure | Risk of disappearance | Probable transformation |
|---|---|---|---|
| Radiologist | Very strong | Low to medium | Validation, integration, interventional |
| Anatomopathologist | Very strong | Low to medium | Augmented digital pathology |
| Dermatologist | Strong | Low | Automated sorting, complex expertise |
| Ophthalmologist | Strong | Low | Delegated screening, specialized treatment |
| General practitioner | Strong | Very low | Increased consultation, coordination |
| Pharmaceutical | Strong | Low | Security, counselling, monitoring |
| Medical biologist | Strong | Low to medium | Validation and integrated interpretation |
| Surgeon | Average | Very low | Robotic assistance and planning |
| Psychiatrist | Average | Very low | Digital monitoring and therapeutic relationship |
| Medical Secretariat | Very strong | Medium to high | Coordination and reception |
| Coding/invoicing | Very strong | High | Exceptional control and audit |
| Clinical research | Strong | Low | Automated recruitment and monitoring |
The fracture line will contrast less man to machine than professionals able to supervise AI to those who use it without critical mind or refuse to integrate it.
Preparing for the future: training, practice and scenarios
Chapter XV How will medical studies evolve?
| Level | Expected competencies |
|---|---|
| 1. General culture | Models, correlation, causality, sensitivity, specificity, bias, drift |
| 2. Clinical use | Tool selection, interpretation, challenge, patient information, traceability |
| 3. Governance | Local validation, audit, security, data protection, incidents, contracts |
| 4. Design | Annotation, clinical criteria, trials, interface between engineers and caregivers |
The training will have to preserve the doctor's intellectual autonomy so that he can challenge the machine, regain his hand and exercise without assistance in a degraded situation.
Chapter XVI What could a medical consultation look like in 2035?
Prior to consultation
- Safe collection of symptoms
- Supplementary matters
- Identification of emergencies
- Summary of the file
- Preparation of assumptions
- Verification of treatments and risks
During consultation
- Ambient transcript
- Information structure
- Suggestion of questions
- Showing recommendations
- Verification of interactions
- Proposal for differential diagnostics
- Preparation of the order
The physician reviews, contextualizes, referees, explains, validates or rejects the proposals.
After consultation
- Proceedings
- Requirements
- Organization of follow-up
- Documentary transmission
- Background
- Remote surveillance
- Alerts in case of worsening
Chapter XVII Three scenarios for tomorrow's medicine
Regulated augmented medicine
The AIA automates repetitive tasks, improves prevention and reduces delays. Sensitive decisions remain supervised. Systems are evaluated, auditable and monitored.
Industrial algorithmic medicine
Productivity requirements dominate. Consultations are standardized and professionals assessed for compliance with automated recommendations. Medicine can become fast but rigid and impersonal.
Two-speed medicine
Affluent patients benefit from an available doctor, human interpretation and personalized follow-up; The others are oriented towards platforms, chatbots and automated protocols.
Chapter XVIII Conditions for responsible integration
1. Prior assessment
- Define clinical need
- Check regulatory status
- Review studies
- Testing the local population
- Assessing organizational impact
- Analyze total cost
2. Local validation
Devices, populations, circuits, practices, data, false positives and atypical cases should be tested locally.
3. Ongoing monitoring
- Measuring performance
- Analyze incidents
- Control drifts
- Compare subpopulations
- Document updates
- Provide for withdrawal
- Organize degraded operation
4. Real human supervision
The professional must be able to understand the tool function, challenge the recommendation, request a second analysis, suspend use and report an incident.
5. Traceability
The file must identify the tool, its version, date, result, human decision, changes and disagreements.
Geography of medical AI: France, Europe, Africa
Chapter XIX Points of vigilance for France and Europe
The assets are a structured system, important foundations, solid clinical research, public infrastructure and regulatory expertise. Fragilities include software fragmentation, insufficient interoperability, heavy procurement, lack of IT resources, heterogeneous data quality, financing and technological dependence.
The priority is not to buy solutions quickly, but to build an institutional capacity to select, test, contract, integrate, monitor, audit and disable tools.
Chapter XX Specific issues for Africa
The continent has a unique terrain: a chronic shortage of medical specialists, insufficient medical demography in the face of population growth, but also a proven ability to rapidly adopt digital technologies when they meet a concrete need — as the distribution of the mobile payment showed.
The AIA can support X-ray reading, ophthalmological screening, teledermatology, emergency screening, community health workers, pregnancy follow-up, translation, epidemiological surveillance, prescription aid and distance training. In areas where access to a specialist may require several hours of travel, a reliable and low-cost screening tool can be a significant health benefit — provided that it is validated on populations and pathologies truly representative of the continent.
Risks include the use of models driven by foreign populations, weak local data, reliance on external suppliers, sensitive data transfers, regulatory insufficiency, instability of infrastructure and lack of maintenance. A dermatological screening algorithm calibrated on mostly clear skins, for example, illustrates in concrete terms the risk of bias already identified in Chapter III: transposed without adaptation, it can produce degraded performance precisely where the clinical need is greatest.
The strategic challenge is to develop local bases, assessment capabilities, research centres, sovereignty rules, balanced partnerships and models that take into account local clinical realities and languages. For health, it is the same challenge of digital sovereignty as that which goes through all African economic sectors in the era of artificial intelligence.
General conclusion Artificial intelligence at the service of care, not in its place
Artificial intelligence will first transform imaging, digital pathology, clinical documentation, surveillance, diagnostic aid, therapeutic customization, pharmaceutical research and hospital organization.
| AIA will accomplish mainly | The professional will retain mainly |
|---|---|
| Detection | Contextualization |
| Calculation | Judgement |
| Sort | Arbitration |
| Executive summary | Verification |
| Monitoring | Liability |
| Documentation | Communication |
| Simulation | Shared decision |
| Standardization | Exception management |
| Statistical prediction | Human interpretation |
The decisive question is not whether the AI will be better than a doctor in an isolated task, but whether a system combining correctly professionals, algorithms, data and organisations will be better than a system based solely on human work.
The most credible future is that of a doctor less mobilized by manual production of information and more responsible for arbitration, supervision, communication, data integration and the guarantee of medical decision-making.
The main threat is not that AI will become a doctor. It is that health systems use it primarily to reduce costs and accelerate flows, without reinvesting gains in human relationship, prevention and quality of care.
References Institutional and scientific sources
- World Health Organization Ethics and governance of artificial intelligence for health: guidance on broad multi-modal models who.int/publications/i/item/9789240084759
- Food and Drug Administration Artificial Intelligence-Enabled Medical Devices fda.gov — Artificial Intelligence-Enabled Medical Devices
- Nature — npj Digital Medicine Systematic review of artificial intelligence in clinical imagining workflows nature.com/articles/s41746-024-01248-9
- NEJM AI Artificial intelligence in radiation oncology a.nejm.org/doi/abs/10.1056/AIra2401164
- NEJM AI Ambient artificial intelligence documentation in clinical practice a.nejm.org/doi/full/10.1056/Aidbp2401267
- European Commission Virtual Human Twins Initiative digital-strategy.ec.europa.eu/en/policies/virtual-human-twins
- WHO Europe Health data governance in the age of artificial intelligence who.int/europe — WHO-EURO-2025-11462-51234-78079
- European Commission Artificial intelligence in healthcare health.ec.europa. had — Artificial intelligence in healthcare
- High Health Authority Digital technologies and professional-oriented d-IA systems has-sante.fr/jcms/p 3363066
- Food and Drug Administration Guidance with digital health content fda.gov — Guidance with digital health content
