CAVaLRi: An Algorithm for Rapid Identification of Diagnostic Germline Variation

Figure 5

CAVaLRi remains equally performant regardless of the source of phenotype terms. The performance of all comparators was evaluated using sets of phenotype terms that were either curated by clinicians (“manual”; solid lines) or generated computationally through NLP processing of medical records (“ClinPhen”; dashed lines) from the test partition for the clinical ES cohort. Accuracy was measured again by (a) PR AUC and (b) top-N curves. CAVaLRi was slightly less accurate in terms of PR AUC and nearly identical in terms of average diagnostic rank when provided with computationally derived phenotype sets. By contrast, the other four algorithms evaluated, hiPHIVE, PhenIX, LIRICAL, and XRare, demonstrated a significant performance decrease when utilizing computationally generated phenotypes.