A real-time AI diagnostic assistant developed by Spanish startup Medsys AI improved physicians' diagnostic accuracy in a controlled simulation study. When using the system, family physicians included the correct diagnosis among their three leading hypotheses in 74.6% of consultations, compared with 62.3% without AI assistance. This represents a relative reduction of 32.6% in diagnostic errors measured by the study.
The findings suggest that AI could support clinical decision-making during consultations. However, the research involved simulated patients rather than actual clinical encounters. Whether the improvements translate into better patient outcomes remains to be established.
AI provides real-time diagnostic suggestions
Unlike diagnostic tools that require physicians to enter specific questions, Medsys AI's Clinical Assistant follows the conversation between doctor and patient. The system transcribes the consultation and uses generative AI to identify potential diagnoses, suggest relevant questions and recommend physical examinations. Its recommendations are supported by information retrieved from more than 1,000 clinical practice guidelines.
The physician remains responsible for interpreting the information and deciding which suggestions to accept or disregard. The study involved 13 Spanish family physicians who completed 260 simulated voice consultations based on 40 diagnostically challenging clinical cases. Consultations were randomly assigned to AI-assisted or unassisted conditions. Three independent physicians evaluated the submitted diagnoses without knowing which consultations involved AI assistance.
The primary outcome was Top-3 diagnostic accuracy: whether the correct diagnosis appeared among up to three diagnoses proposed by the physician. With AI assistance, adjusted accuracy increased by 12.3 percentage points. The improvement was statistically significant. However, the study found no statistically conclusive improvement in Top-1 accuracy, which measures whether the physician's first diagnosis was correct. Consultations using AI also took approximately 10.7% longer on average, suggesting that improved diagnostic reasoning does not necessarily translate into immediate time savings.
Safety questions remain
The researchers also examined whether physicians might become overly dependent on incorrect AI recommendations. When the system presented an incorrect diagnostic suggestion, physicians sometimes incorporated that diagnosis into their final assessment. The study identified a possible overreliance signal, although the difference between assisted and unassisted conditions was not statistically significant.
This highlights an important challenge for clinical AI: improving diagnostic performance without encouraging automation bias. The researchers emphasise that the study was designed as a simulation-based stress test rather than an assessment of routine clinical effectiveness. The company funded the research, and founder Manuel Pinilla participated in the study and is listed as a co-author. Diagnostic assessments were performed by independent evaluators. Medsys AI is now conducting further clinical evaluation at a hospital in Madrid. The company is also preparing additional fundraising to support regulatory approval, research and potential commercialisation.
AI increasingly supports diagnostic reasoning
The findings add to growing evidence that AI can assist physicians with complex diagnostic decisions. Earlier this year, ICT&health Global reported on DeepRare, an AI system developed to identify rare diseases by combining genomic information, medical records and specialised diagnostic tools. In a comparative evaluation involving 163 complex cases, DeepRare achieved a first-attempt diagnostic accuracy of 64.4%, compared with 54.6% among experienced physicians.
The two studies investigate different applications and cannot be directly compared. DeepRare focuses on complex rare disease diagnosis, while Medsys AI provides real-time support during primary care consultations. Nevertheless, both demonstrate the potential of AI to identify diagnostic possibilities that clinicians might otherwise overlook.
Reduce unnecessary diagnostic testing
Medsys AI also expects its technology to help reduce unnecessary diagnostic testing. The company estimates that a hypothetical 10% reduction in emergency department tests could generate annual savings of €210 million across Spain's healthcare system. These figures are company projections, not demonstrated savings. The current study did not measure test utilisation, healthcare expenditure or patient outcomes.
Further prospective research will therefore be necessary to determine whether improved diagnostic accuracy translates into safer care, fewer unnecessary investigations and more efficient clinical workflows. For healthcare organisations, the central question is no longer simply whether AI can generate useful diagnostic suggestions, but whether integrating those suggestions into everyday clinical practice produces measurable benefits without compromising professional judgement or patient safety.
References
Add ICT&health on Google
Show more content from ICT&health in Google Search.
