Feature Mixing Approach for Detecting Intraoperative Adverse Events in Laparoscopic Roux-en-Y Gastric Bypass Surgery

TL;DR

This work tackles rare intraoperative adverse events during Roux-en-Y gastric bypass by introducing BetaMixer, a transformer-based framework that jointly classifies IAEs and regresses their severity on a $0$-$5$ scale. It leverages Beta distribution-based sampling to transform discrete severity labels into continuous targets and uses a generator-discriminator pair to normalize features, followed by a temporal transformer with regression tokens for per-event prediction. On the extended MultiBypass140 dataset, BetaMixer achieves state-of-the-art metrics (e.g., weighted F1 $=0.76$, recall $=0.81$, PPV $=0.73$, NPV $=0.84$) and demonstrates the value of short temporal context (5-frame inputs) for accurate detection and severity estimation. The approach offers robust, real-time IAE detection and quantification under imbalanced data, with implications for improved surgical safety and decision support.

Abstract

Intraoperative adverse events (IAEs), such as bleeding or thermal injury, can lead to severe postoperative complications if undetected. However, their rarity results in highly imbalanced datasets, posing challenges for AI-based detection and severity quantification. We propose BetaMixer, a novel deep learning model that addresses these challenges through a Beta distribution-based mixing approach, converting discrete IAE severity scores into continuous values for precise severity regression (0-5 scale). BetaMixer employs Beta distribution-based sampling to enhance underrepresented classes and regularizes intermediate embeddings to maintain a structured feature space. A generative approach aligns the feature space with sampled IAE severity, enabling robust classification and severity regression via a transformer. Evaluated on the MultiBypass140 dataset, which we extended with IAE labels, BetaMixer achieves a weighted F1 score of 0.76, recall of 0.81, PPV of 0.73, and NPV of 0.84, demonstrating strong performance on imbalanced data. By integrating Beta distribution-based sampling, feature mixing, and generative modeling, BetaMixer offers a robust solution for IAE detection and quantification in clinical settings.

Figures (7)

• Figure 1: Illustration of an occurrence of intraoperative adverse event---a case of bleeding---showing different levels of severity.
• Figure 2: IAE frequency per (a) phase and (b) step across the two data centers.
• Figure 3: Beta distribution sampling of the grades of adverse events in (top) training and (bottom) testing sets.
• Figure 4: Overview of BetaMixer: The backbone $\mathcal{B}$ extracts features, which are transformed into a normal distribution by the generator $\mathcal{H}$. A transformer with positional embeddings encodes, classifies, and regresses IAE severity, while the discriminator $\mathcal{D}$ ensures feature normalization.
• Figure 5: Architecture of the transformer model - illustrating self-attention computation among frame features and independent cross-attention with the query token for a given IAE. The corresponding output is regressed to predict the severity.