JMIR Publications

ABSTRACT

Background:

Polypropylene mesh was extensively utilized in surgical procedures to treat pelvic organ prolapse (POP) due to its cost-efficiency and durability. However, complications, including mesh exposure through the vaginal wall, prompted discontinuation of its usage for POP treatment in some countries. Developing predictive models via supervised machine learning holds promise in identifying risk factors associated with such complications, thereby facilitating better informed surgical decisions. Previous studies have demonstrated the efficacy of anticipating medical outcomes by employing supervised machine learning approaches that integrate patient healthcare data with laboratory findings. However, such approach has not been adopted within the realm of POP mesh surgery.

Objective:

We examined the efficacy of supervised machine learning to predict POP post-surgical mesh exposure utilizing three different datasets: (1) patient medical record data, (2) biomaterial-induced blood cytokine levels, and (3) the integration of both.

Methods:

Blood samples and healthcare data were collected from 20 female patients who had prior surgical intervention for POP using transvaginal polypropylene mesh. Of these subjects, 10 had experienced mesh exposure through the vaginal wall following surgery and 10 had not. Standardized medical record data, including vital signs, previous diagnoses, and social history, was acquired from patient records. Cytokine levels in plasma obtained from patient blood samples incubated with sterile polypropylene mesh were measured via multiplex assay. Datasets were created with patient medical record data alone, blood cytokine levels alone, and (3) the integration of both. 70% and 30% of the data were split into training and testing sets, respectively, for machine learning models that predicted the presence or absence of post-surgical mesh exposure.

Results:

Upon training the models with patient medical record data, systolic BP, pulse pressure, and a history of alcohol usage emerged as the most significant factors for predicting mesh exposure. Conversely, when the models were trained solely on blood cytokine levels, IL-1β and IL 12 p40 stood out as the most influential cytokines in predicting mesh exposure. Using the combined dataset, new factors emerged as the primary predictors of mesh exposure: IL-8, TNF-α, and the presence of hemorrhoids. Remarkably, models trained on the integrated dataset demonstrated superior predictive capabilities with a prediction accuracy as high as 94%, surpassing the predictive performance of individual datasets.

Conclusions:

Supervised machine learning models demonstrated improved prediction accuracy when trained using a composite dataset comprising patient medical record data and biomaterial-induced blood cytokine levels, surpassing the performance of models trained with either dataset in isolation. This result underscores the advantage of integrating healthcare data with blood biomarkers, presenting a promising avenue for predicting surgical outcomes in not only POP mesh procedures but also other surgeries involving biomaterials and enhancing informed decision-making for both patients and surgeons, ultimately elevating the standard of patient care.