Optimal Single or Combined Parameters for Dyssynergic Defecation on Anorectal Manometry: A Proof-of-Concept Machine...

To perform a proof-of-concept study to evaluate the performance of individual and combined anorectal manometry and balloon expulsion testing (BET) parameters for diagnosing dyssynergic defecation (DD) using machine learning (ML).

A single gastrointestinal motility specialist assessed (2008-2018) 307 patients. A ML model was constructed on anorectal manometry/BET data. BET was weight-based in 235 patients and time-based in 72 patients. Missing data were imputed. Data were split 75% train/25% test with preprocessing performed. Five common ML models were trained, and performance was compared using receiver operating characteristic (ROC) area under the curve and accuracy based on 100 bootstrapped samples. The most accurate models were fine-tuned and evaluated on the test set.

The median age was 40 years, body mass index 22.2 kg/m 2 , and 81% were women. Two-hundred and 22 (72.3%) patients had a clinical diagnosis of DD. Compared with patients without clinical diagnosis of DD, patients with DD had significantly lower rectoanal pressure gradient (RAG) (median -34.7 [IQR -57.1, -13.0] vs -24.6 [IQR -44.5, -2.4] mm Hg, P = 0.009) and more abnormal BET (59.5% vs 11.8%, P < 0.001). Four features were retained in the optimized model as predictors of DD: abnormal BET, greater resting anal pressure (RAP), and more negative RAG. The optimal parameter was abnormal BET ( P = 0.003), but combinations of 2 manometry results (RAP and RAG) yielded positive predictive values (PPVs) >80%. The optimized logistic regression model had an area under the curve of 0.878, with a sensitivity of 75.0% and specificity of 81.8% at a probability threshold of 0.704.

Abnormal BET, or a combination of 2 manometric parameters (RAP and RAG), provides >80% positive predictive value for diagnosing DD in patients with chronic constipation.