This paper by Kauffman et al. introduces InfEHR, a framework that applies deep geometric learning to electronic health records (EHRs) to improve clinical phenotype detection. Unlike conventional models that depend on large, labeled datasets, InfEHR works effectively with minimal labeling by converting whole EHRs into temporal graphs that capture patient health trajectories. These graphs allow the system to compute and revise clinical likelihoods automatically, which allows clinicians to have unbiased and dynamic representations of patient phenotypes. The framework was tested using data from the Mount Sinai Health System and UC Irvine Medical Center on two conditions: neonatal culture-negative sepsis (3% prevalence) and postoperative acute kidney injury (21% prevalence). InfEHR was compared against physician heuristics, which are typically used when labeled data are sparse. In both conditions, InfEHR demonstrated superior performance. For culture-negative sepsis, sensitivity improved dramatically (0.60 vs. 0.04) while maintaining high specificity (0.98 vs. 0.99). For acute kidney injury, sensitivity was also much higher (0.71 vs. 0.20) with strong specificity (0.93 vs. 0.98). These findings show that InfEHR is particularly well-suited for detecting low-prevalence diseases where traditional approaches underperform.