In a recent Viewpoint published in The Lancet Digital Health we evaluate various AI performance measures essential for supporting medical decisions. It highlights the significance of selecting appropriate metrics to avoid misleading outcomes that could adversely affect patient care and incur additional costs. We determined 32 performance measures across five domains (discrimination, calibration, overall performance, classification, and clinical utility). Using the ADNEX model, which predicts ovarian tumor malignancy, as a case study, we determined that performance assessment should prioritize discrimination, calibration, and clinical utility to improve decision-making in medical practice. Read the article here.

In research led by Sara Keller, a visiting scholar in the Boussard lab,  a predictive model for acute care use (ACU) in oncology patients receiving systemic therapy, was used to quantify costs. Utilizing Medicare claims data from 20,556 cancer patients treated at an academic medical center between 2010 and 2022, the analysis revealed that 18.58% of the patients experienced at least one ACU event, with average daily costs significantly higher for those affected (US $94.62) compared to those without ACU events (US $53.28). The predictive model demonstrated substantial financial impact, yielding projected savings of US $910,000 in the first year and increasing to US $9.46 million by year six, totaling approximately US $31.11 million over six years, based on an assumed 35% prevention rate for ACU events. These findings underscore the potential of predictive analytics to reduce costs associated with ACU in oncology, enhancing economic efficiency in cancer care while highlighting the need for further research into the associated health benefits. For more details read the research here.

Hidradenitis suppurativa (HS) is a chronic inflammatory skin condition associated with elevated rates of comorbid mental health disorders.  Dr. Leandra A. Barnes's lab studied the association between biologic treatment for this condition and the risk of mental health diagnoses. A retrospective cohort study was conducted using de-identified electronic health records from the TriNetX network, which includes adults diagnosed with HS, with findings that suggest a potential benefit of biologics in improving mental health outcomes. This study highlights the importance of integrated care addressing both physical and mental health in HS management, emphasizing the role of biologic therapy in alleviating psychiatric comorbidities through effective disease control and reduced psychosocial stressors. For more details, read the study in the Journal of Investigative Dermatology here.

In the evolving landscape of AI, the belief that more data automatically produce better models has fueled a rapid and often unchecked reliance on synthetic data to expand training datasets. Although synthetic data can help address real shortages in high quality clinical data, they often fall short in representing the nuances that matter most in medicine. Synthetic data aim to mirror population level distributions, yet they frequently fail to capture rare conditions, intersectional identities, and complex comorbidities. Even when they comply with privacy regulations such as HIPAA and GDPR, these limitations render synthetic data unfit for high stakes clinical modelling where patient level accuracy is non negotiable. A growing concern is the emergence of synthetic trust, where confidence in a model is based on artificially generated data that fail to reflect clinical validity or demographic reality. In our Viewpoint published in Lancet Digital Health, we call for greater caution in the use of synthetic data for clinical algorithms. We propose a set of actionable safeguards, including minimum standards for training data, fragility testing during model development, and clear disclosures when synthetic data contribute to model outputs. Together, these measures aim to strengthen accountability and preserve data integrity and fairness when synthetic data are used in health care.