The volume of research on artificial intelligence (AI) in medicine has grown sharply. According to a bibliometric review published as a preprint on medRxiv (Awasthi et al., 2024), the number of articles on AI and machine learning in healthcare indexed on PubMed rose from 3,351 in 2019 to 23,306 in 2023.
Yet the gap between published research and clinical practice remains stark. Van de Sande et al. (npj Digital Medicine, 2024) found that fewer than 2% of AI models developed in research reach stable clinical application. The obstacle is not purely technical: it concerns the quality of the data on which models are trained, and their compatibility with regulatory standards. That analysis was the starting point for S-RACE.
Professor Carlo Tacchetti and Professor Antonio Esposito, coordinators of the S-RACE centre established at Vita-Salute San Raffaele University (UniSR) and IRCCS Ospedale San Raffaele with the support of Microsoft, responded to this challenge with the S-RACE platform, an AI system built to develop prognosis and treatment models in medicine.
Before designing S-RACE, the researchers conducted a systematic review of the scientific literature to identify why existing AI models fail in practice. Two main causes emerged. «The first concerns the quality and biases of the data on which most models are trained. A model trained on data collected inconsistently, or according to criteria that vary from one institution to another, produces unreliable predictions the moment the context changes» Tacchetti explains.
The second is regulatory compliance: a model that does not meet European standards such as the EU AI Act, the GDPR, ISO 42001 or the Risk Management Framework of the US National Institute of Standards and Technology cannot be used in clinical settings, regardless of its technical performance. These two constraints shaped the architecture of the S-RACE platform.
The S-RACE platform is cloud-hosted and structured in multiple stacked layers. The first layer concerns data acquisition: in full compliance with privacy regulations, the platform accesses hospital databases and imports records for patients selected for a study designed around a specific scientific question. For each patient, S-RACE imports clinical records, radiology reports and other documents, extracts the relevant data and classifies it according to the UMLS (Unified Medical Language System), an international medical vocabulary for standardised communication of health data.
The extracted data are then encoded according to the FHIR (Fast Healthcare Interoperability Resources) protocol, an international standard for representing and exchanging health information, and subjected to pseudonymisation before moving to the analysis layer. Pseudonymisation replaces personal identifiers such as names and other direct attributes with aliases or indirect identifiers, making the data non-attributable to the individual.
The second layer covers the analysis of the extracted, pseudonymised data across four sequential phases. The first two assess whether the collected data are adequate to answer the initial research question: data quality evaluation and exploratory data analysis, the latter identifying patients for whom certain data points are missing.
This curated, standardised archive of real clinical data forms the foundation on which any AI model in medicine is trained. Each model is released only after external validation on patient cohorts from hospitals that did not participate in the original study, a condition that verifies the model is not overfitted to San Raffaele Hospital’s own data.
One of the case studies completed using S-RACE focuses on kidney cancer, developed in collaboration with Professor Andrea Salonia and and Dr Alessandro Larcher of the Urological Research Institute (URI) at San Raffaele Hospital. The clinical question was: can we predict, before surgery, which patients with renal carcinoma will develop a recurrence?
The question has a direct clinical rationale. Adjuvant chemotherapy, administered after surgery, is given to only around 30% of patients who develop recurrence. «Knowing before surgery how the tumour is likely to evolve, whether it will recur and with what probability, would help clinicians determine whether a patient might benefit from neoadjuvant chemotherapy, administered before the operation to reduce tumour mass and lower the risk of recurrence» Tacchetti explains.
In kidney cancer, however, a biopsy before surgery is often not feasible because of the invasiveness of the procedure. Without this histopathological data, studies attempting to evaluate the efficacy of neoadjuvant chemotherapy have not produced conclusive findings.
Tacchetti and colleagues used S-RACE to extract from the data of approximately 3,000 patients with renal carcinoma a set of parameters relevant to recurrence risk. «In this case, starting from real-world data, S-RACE extracted the most informative features to build a prognosis model for the tumour» he adds. The study also compared the platform’s automated pathway with clinicians’ manual variable selection.
Clinicians had manually selected 69 predictive variables for recurrence risk; the platform identified 206. Of the eight parameters that proved relevant in the final model, six were common to both pathways. Two, however, haemoglobin levels and platelet counts, had not surfaced through manual analysis but emerged through S-RACE: parameters no one had previously considered relevant to kidney cancer prognosis, yet fundamental.
One of the most critical issues in applying AI in medicine is model interpretability: the ability to make transparent to the clinician the reasoning behind an algorithmic prediction. This is one of the main reasons why so-called “black box” models, despite often being the highest-performing, rarely pass clinical scrutiny. «Responsibility for the diagnostic, prognostic and therapeutic decision rests with the physician» Tacchetti emphasises. «If I use an algorithm to support my decision but cannot see its reasoning, I cannot make an informed judgement».
S-RACE therefore adopts a "white box" approach and integrates techniques such as SHAP (SHapley Additive exPlanations), which quantify the contribution of each parameter to the model’s prediction. In the kidney cancer example, the platform dashboard shows, alongside the calculated recurrence risk for each patient, the relative weight of each variable in defining that risk and guidance on interpreting it. This is complemented by error analysis: a system that identifies, among the patients under examination, those for whom the model is most likely to produce an incorrect prediction.
Among S-RACE’s most significant innovations is its federated architecture. Each hospital participating in a multicentre study does not transfer its data to a central server: instead, the platform exports computational resources to each participating centre. Every centre trains its model locally; the models are then centralised, aggregated into a single model, and redistributed to the centres in an iterative cycle. «In this way, participating centres do not have to share their data, preserving confidentiality» Tacchetti explains.
Equally central is the question of real-world data: data collected in routine clinical practice, on unselected populations. Unlike classical randomised controlled trials, which validate a drug on highly homogeneous and carefully selected patient cohorts, real world data reflect genuine clinical complexity: comorbidities, concurrent therapies, variability in the timing of examinations, differences in follow-up. Working on this data makes it possible to develop predictive models that are more representative of the patients a clinician will encounter.
As of May 2026, S-RACE hosted 22 active research projects, covering areas including diabetes, oncology, cardiovascular diseases, neurology and neurocritical care, with data from approximately 65,000 patients. The transition to genuine clinical use, in which the model informs a real-time decision for an individual patient, remains to be made. Tacchetti is candid about the pace of change in the field.
«AI is evolving so rapidly that if you had asked me how I see the platform by the end of this year, I would not have had time to answer», he recalls. What stays constant, he argues, is what the next generation of physicians needs to do: «these systems will become increasingly pervasive in healthcare. Doctors need to start studying them now and keep up continuously with how the technology evolves, so they can use it responsibly».
Artificial Intelligence in Healthcare: 2023 Year in Review, Awasthi et al., 2024, medRxiv
To warrant clinical adoption AI models require a multi-faceted implementation evaluation, van de Sande et al., 2024, npj digital medicine
Regolamento (UE) 2024/1689 del Parlamento europeo e del Consiglio, del 13 giugno 2024