Assessing Patient Eligibility for Inspire Therapy through Machine Learning and Deep Learning Models

Mohsena Chowdhury, Tejas Vyas, Rahul Alapati, Andrés M. Bur, Guanghui Wang

Introduction

Obstructive Sleep Apnea (OSA) is a prevalent condition characterized by the repetitive collapse of the upper airway during sleep, leading to fragmented rest and significant cardiovascular risks. While Continuous Positive Airway Pressure (CPAP) remains the gold standard for treatment, patient adherence is notoriously low due to discomfort and lifestyle constraints. For these patients, Inspire therapy—an FDA-approved implantable neurostimulator that targets the hypoglossal nerve to maintain airway patency—offers a transformative alternative.

However, the success of Inspire therapy is not universal. Clinical eligibility currently relies on Drug-Induced Sleep Endoscopy (DISE), a procedure where surgeons manually evaluate airway collapse patterns while the patient is sedated. This manual assessment, often recorded as a "VOTE" score, is subjective and prone to inter-observer variability. To address this, researchers have developed the first machine learning framework designed to automate the screening process, aiming to improve patient selection and prevent unnecessary surgical interventions for those unlikely to respond.

Methodology

The study implemented a dual-stream classification strategy to evaluate patient candidacy from both visual and clinical perspectives:

1. Deep Learning for Video Analysis

The researchers focused on two critical anatomical regions captured during DISE: the Velopharynx (VP) and the Base of the Tongue (BOT). They benchmarked six established Convolutional Neural Network (CNN) architectures to classify frames as indicative of a "Responder" or "Non-responder":

  • VGG-16
  • ResNet-50 and ResNet-101
  • EfficientNet-B0
  • DenseNet-121 and DenseNet-169

To bridge the gap between individual image classification and clinical decision-making, the team employed a patient-level majority-voting scheme. This approach aggregates predictions across a sequence of video frames to produce a single classification for the patient, smoothing out noise from inconsistent lighting or camera angles.

2. Machine Learning for Clinical Data

In parallel, the study utilized 22 clinical features, including BMI, demographic data, and pre-operative apnea severity. Five traditional machine learning algorithms were trained on this data:

  • Logistic Regression
  • Decision Trees
  • Random Forest
  • k-Nearest Neighbors
  • Gradient Boosting

Dataset and Evaluation

The models were trained and validated using data from 127 patients at the University of Kansas Medical Center. The dataset comprised:

  • Video Data: 24,750 annotated frames (11,298 for VP; 13,479 for BOT).
  • Clinical Data: 22 features per patient.
  • Ground Truth: Defined by the Sher Criteria, requiring a >50% reduction in post-operative apnea events.

To ensure the models generalized well to new patients, the researchers utilized 10-fold patient-wise cross-validation, preventing the models from "memorizing" specific patient characteristics.

Key Findings

The Significance of the Velopharynx

The study revealed that the Velopharynx (VP) is a more reliable predictor of surgical success than the Base of the Tongue (BOT). Models trained on VP data consistently outperformed those trained on BOT data, suggesting that collapse patterns in the upper throat are more discriminative for neurostimulation outcomes.

Performance Metrics

  • Top Performer: DenseNet-169 achieved the highest patient-level accuracy at 71.3% when analyzing the Velopharynx.
  • Clinical Prediction: Interestingly, Logistic Regression using only clinical data achieved a respectable 68.5% accuracy. This indicates that while visual data is crucial, traditional metrics like BMI and baseline severity remain highly informative.
  • Precision and Safety: DenseNet-169 demonstrated a precision rate of approximately 78%. In a clinical context, high precision is vital as it minimizes "false positives," ensuring that patients are not subjected to an invasive surgery that will not benefit them.

Temporal Consensus

The transition from frame-level to patient-level analysis proved essential. For instance, the VGG-16 model saw its accuracy increase from 67.6% to 71.1% simply by implementing the majority-voting consensus, highlighting the importance of temporal context in medical video analysis.

Limitations and Future Directions

While this study represents a significant step toward automated OSA diagnostics, the sample size of 127 patients is relatively small for deep learning applications. This likely accounts for the accuracy ceiling of approximately 71%. Furthermore, the current models analyze video and clinical data in isolation.

The research team is now moving toward a multimodal fusion framework. By integrating visual features from DISE videos with clinical text data into a unified model, they aim to capture a more holistic physiological profile. Such an integrated system could provide the high-confidence predictions necessary for routine clinical deployment, ultimately optimizing outcomes for sleep apnea patients worldwide.

Summary updated

BibTeX
@inproceedings{chowdhury2024assessing,
  author = {Mohsena Chowdhury and Tejas Vyas and Rahul Alapati and Andrés M. Bur and Guanghui Wang},
  title = {{Assessing Patient Eligibility for Inspire Therapy through Machine Learning and Deep Learning Models}},
  booktitle = {ICBBT},
  year = {2024},
  doi = {10.1145/3674658.3674687},
}
Figure