Rome, November 9 (aaskanews) – Italian spending on digital healthcare will reach €2,7 billion in 2025, a 9 percent increase, according to the Digital Healthcare Observatory at the Polytechnic University of Milan. Seventy-one percent of healthcare facilities list artificial intelligence among their strategic areas, while diagnostic support solutions are included in 11 percent.
This is where predictive models from clinical research come in, such as a study by the University of Siena on lithium in bipolar disorder. Among specialists and general practitioners, 61 percent have used generative artificial intelligence tools in the past year, almost always generalist platforms. The study, "Utilizing Machine Learning to Forecast 3-Month Remission Outcomes in Bipolar Disorder Patients Treated with Lithium," conducted by Professor Andrea Fagiolini's group, examines whether data available at the start of lithium therapy allow for estimating the remission of depressive symptoms after three months. Alessandro Cuomo, associate professor of Psychiatry at the University of Siena, notes that international guidelines continue to recommend lithium as a first-line option, especially for maintenance and acute mania. The challenge lies in predicting response.
Two people with the same diagnosis can have very different outcomes, symptoms, medical conditions, and therapeutic responses, explains Cuomo. Hence the study's hypothesis: Machine learning attempts to verify whether their interactions offer a more accurate estimate than that obtained by looking at one variable at a time. According to the authors, the retrospective analysis involved 593 patients, with sociodemographic, clinical, laboratory data, concomitant therapies, and symptom scales.
Random forest, XGBoost, neural network, and support vector machine were compared, with 80 percent of the sample used for training and 20 percent for testing. Internal validation remains: all patients come from the same cohort, and the model still needs to be demonstrated in people recruited elsewhere and at a later time point, Cuomo specifies. Remission, defined as a score of up to 10 on the Montgomery sberg Depression Rating Scale, was achieved by 44 percent of patients. The best model, a random forest with polynomial transformations, achieved an AUC of 0,76 and an accuracy of 0,64. Cuomo explains the first indicator: by randomly choosing one person in remission and one not in remission, the model assigns a higher score to the former in approximately 76 out of 100 comparisons. Regarding the threshold, he adds that this threshold measures depressive symptoms and, by itself, does not demonstrate global euthymia, functional recovery, or long-term stability. Among the most informative variables are the initial severity of depressive and manic symptoms, creatinine, TSH, body mass index, and age. Cuomo introduces a distinction: A variable can help the algorithm sort the probabilities without being the cause of the outcome. The abstract, he reports, does not publish the direction of the association for each variable. Regarding clinical use, the position is clear. Today, the decision remains entrusted to specialist evaluation, Cuomo states. Monitoring remains necessary: Lithium has a narrow therapeutic range, he recalls, and requires checks of lithium levels and renal and thyroid function.
Askanews