Improving Pelvic Floor Muscle Training with AI: A Novel Quality Assessment System for Pelvic Floor Dysfunction

Batoul El‐Sayegh, Chantale Dumoulin, François Leduc-Primeau, Mohamad SawanSensors 2024 · DOI

TL;DR

This paper introduces an AI-powered system to automatically assess the quality of pelvic floor muscle (PFM) contractions, a crucial component of rehabilitation for urinary incontinence. By combining a deep learning-based detector with a feature-extraction classifier, the system provides objective feedback, potentially revolutionizing how women train and improve their PFM function at home and in clinical settings.

What Problem Does This Paper Address?

Urinary incontinence (UI) significantly impacts women's quality of life, and pelvic floor muscle training (PFMT) is the primary treatment. However, many women struggle to perform PFM contractions correctly due to a lack of awareness or misconceptions about the movement. Current assessment methods, like digital palpation by a physiotherapist, are subjective, time-consuming, and not feasible for remote or home-based training. This creates a critical need for an objective, automated system that can accurately assess PFM contraction quality and provide actionable feedback.

Methods

The proposed system has two main components:

  1. PFM Contraction Detector: This component uses a Convolutional Neural Network (CNN), specifically a 1D stacked CNN (1DSCNN), to distinguish between actual PFM contractions and non-contractions. The CNN was trained on time-series force data from vaginal dynamometry.
  2. PFM Contraction Performance Classifier: This component autonomously quantifies the quality of maximal PFM contractions. It employs a custom feature extraction algorithm that models the contraction graph as a piecewise linear trapezoid. Key features like maximum force, contraction time, slope, and relaxation are extracted. These features are then fed into a Random Forest Classifier (RFC) to predict the strength rating based on the Modified Oxford Scale (MOS).

Datasets

The study utilized two datasets:

  • Dataset 1: Acquired from a previous randomized non-inferiority trial using an intra-vaginal dynamometer. This dataset included PFM contraction, rest, and cough/push recordings from 318 participants. This dataset was used for training and validation of the CNN and for developing the PFM strength classifier.
  • Dataset 2: Obtained from a feasibility study using a newly developed portable PFM dynamometer. This dataset, from 13 participants, was used to test the generalization ability of the CNN model.

Key Insights and Ideas

  • AI for Objective PFM Assessment: The paper demonstrates that AI, particularly CNNs and RFCs, can effectively automate the assessment of PFM contractions, moving beyond subjective manual evaluations.
  • Bridging the Gap in PFM Training: The system addresses the critical need for objective, real-time feedback, empowering individuals to self-monitor and improve their PFM training, which is essential for effective PFMT.
  • Feature Engineering for Quality: The novel feature extraction algorithm, based on a trapezoidal model of contraction graphs, successfully captures clinically relevant aspects of PFM function.
  • Generalizability is Key: Testing the CNN on a different dynamometer dataset highlights the system's potential to be device-agnostic, a crucial factor for widespread adoption.

Key Results

  • PFM Contraction Detection: The CNN-based detector achieved 97% accuracy on Dataset 1 and 100% accuracy on Dataset 2, demonstrating high performance in distinguishing contractions from non-contractions.
  • PFM Strength Classification: The RFC classifier achieved 95.56% accuracy in classifying PFM strength within a ±1 MOS class, outperforming other tested methods like XGB and LR.
  • Feature Extraction Accuracy: The extracted features showed a Root Mean Squared Percentage Error (RMSPE) below 5% compared to manually identified values, indicating high accuracy in feature extraction.
  • Overall System Performance: The system demonstrated strong performance in both detection and classification, suggesting its potential for clinical utility.

Limitations and Future Work

  • Task Dependency: The current system's evaluation is dependent on specific task instructions (rest, contraction, push, cough). Future work should explore more standardized task instructions to reduce variability.
  • Proxy for Non-Contraction: The study used rest, push, and cough tasks as proxies for non-PFM contraction. Future research should aim to test with data recorded directly from incorrect contractions for better validation.
  • Dataset Bias: Dataset 1 had a bias towards higher MOS scores, as participants were required to contract their PFMs. Future studies should include a larger population with a wider range of PFM strengths (MOS 0-5) and consider patients without prior PFM pain.
  • Integration and Refinement: The authors plan to integrate the quality assessment system into a novel portable PFM dynamometer. Further refinement of the overall rating scale, including an ablation study to assess feature significance and weights, is also planned.
  • Diverse Environments: Further evaluation is needed to determine the model's robustness in more diverse and less-controlled environments.

Summary updated

Abstract

The first line of treatment for urinary incontinence is pelvic floor muscle (PFM) training, aimed at reducing leakage episodes by strengthening these muscles. However, many women struggle with performing correct PFM contractions or have misconceptions about their contractions. To address this issue, we present a novel PFM contraction quality assessment system. This system combines a PFM contraction detector with a maximal PFM contraction performance classifier. The contraction detector first identifies whether or not a PFM contraction was performed. Then, the contraction classifier autonomously quantifies the quality of maximal PFM contractions across different features, which are also combined into an overall rating. Both algorithms are based on artificial intelligence (AI) methods. The detector relies on a convolutional neural network, while the contraction classifier uses a custom feature extractor followed by a random forest classifier to predict the strength rating based on the modified Oxford scale. The AI algorithms were trained and tested using datasets measured by vaginal dynamometry, combined in some cases with digital assessment results from expert physiotherapists. The contraction detector was trained on one dataset and then tested on two datasets measured with different dynamometers, achieving 97% accuracy on the first dataset and 100% accuracy on the second. For the contraction performance classifier, the results demonstrate that important clinical features can be extracted automatically with an acceptable error. Furthermore, the contraction classifier is able to predict the strength rating within a ±1 scale point with 97% accuracy. These results demonstrate the system's potential to enhance PFM training and rehabilitation by enabling women to monitor and improve their PFM contractions autonomously.
BibTeX
@inproceedings{elsayegh2024improving,
  author = {Batoul El‐Sayegh and Chantale Dumoulin and François Leduc-Primeau and Mohamad Sawan},
  title = {{Improving Pelvic Floor Muscle Training with AI: A Novel Quality Assessment System for Pelvic Floor Dysfunction}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24216937},
}
Figure