Model generalizability considerations in development and evaluation of SaMD

Software as a Medical Device (SaMD) has been transforming medical practices by improving patient care with more precise and timely information. Key capabilities of SaMD are often powered by Artificial Intelligence (AI) algorithms. However, significant challenges arise in developing robust algorithms for SaMD, as these algorithms can perform well during development but poorly during pivotal validation studies. Due to the rapid advancement in medical research, data from new studies or real-world sources are likely to differ significantly from the legacy data used for development; with this, algorithms need to account for these potential data heterogeneity differences. This paper discusses the shortcomings of conventional cross-validation methods widely used in SaMD algorithm development and demonstrates model performance overestimation in the presence of data heterogeneity. To address this bias, we propose a simple and practical alternative: the leave-one-set-out (LOSO) cross-validation method. Additionally, we outline best practices for designing independent validation pivotal studies.