Prototype Learning for Medical Time Series Classification via Human-Machine Collaboration

Jia Xie, Zhu Wang, Zhiwen Yu, Yasan Ding, Bin GuoSensors 2024 · OpenReview

TL;DR

This paper introduces PahNet, a novel deep learning model for medical time series classification (specifically ECG signals) that combines prototype learning with attention mechanisms. It addresses the critical need for interpretable AI in healthcare by allowing domain experts to collaboratively refine the model's learned "prototypes," leading to more trustworthy and accurate diagnoses.

What Problem Does This Paper Address?

Deep learning models excel at complex pattern recognition, but their "black box" nature makes it difficult to understand why they make certain predictions. This lack of interpretability is a major hurdle for their adoption in high-stakes fields like healthcare, where trust and validation by domain experts (like physicians) are paramount. Traditional time series analysis methods often struggle with capturing complex dependencies or providing clear explanations. Existing interpretable AI methods, like attention mechanisms, can be opaque to non-experts, and prototype learning, while intuitive, can generate too many prototypes, obscuring the decision-making process. This paper aims to bridge this gap by creating a model that is both accurate and inherently interpretable, fostering collaboration between AI and human experts.

Methods

The core of the proposed method, PahNet, is a deep sequence network that integrates three key components:

  • Sequence Encoder (L): This component, using a Bidirectional Long Short-Term Memory (Bi-LSTM) network with an attention mechanism, processes the input time series (ECG segments) to extract high-quality, informative features. The attention mechanism helps focus on the most crucial parts of the signal.
  • Prototype Learning Layer (P): This layer learns a set of "prototypes" – representative examples of different classes (e.g., normal heartbeats, atrial fibrillation). The model learns to classify new inputs by comparing them to these prototypes.
  • Fully Connected Layer (F): This final layer takes the similarity scores between the input and prototypes to produce the classification output.

A novel aspect is the human-machine collaboration mechanism. Instead of just a "human-in-the-loop" where humans correct errors, PahNet allows domain experts to directly refine the learned prototypes. This means physicians can adjust these representative examples based on their expertise, making the model's decisions more aligned with clinical reality. The paper also introduces a prototype quantity control method to manage the number of prototypes, preventing them from becoming overwhelming and ensuring clarity.

Datasets

The primary dataset used in this study is the PhysioNet/Computing in Cardiology (CinC) Challenge 2017 dataset. This dataset contains 8528 single-lead ECG recordings categorized into four classes: normal sinus rhythm, atrial fibrillation (AF), other rhythm, and noisy recordings. The paper focuses on classifying the first three categories.

Key Insights and Ideas

  • Prototypes as Intuitive Explanations: The paper highlights that learned prototypes, when visualized and interpreted, can serve as intuitive explanations for model decisions, making them accessible to domain experts.
  • Human-Machine Collaboration is Key: The authors argue that true interpretability and trustworthiness in medical AI come from a collaborative process where human expertise actively shapes the model, not just passively corrects it.
  • Balancing Accuracy and Interpretability: PahNet demonstrates that it's possible to achieve competitive accuracy while significantly improving interpretability, a crucial trade-off in medical AI.
  • Attention Enhances Prototype Quality: The integration of attention mechanisms within the Bi-LSTM encoder helps generate more refined and informative features, leading to better prototype learning.

Key Results

The PahNet model was evaluated on ECG classification tasks:

  • Binary Classification (Normal Sinus Rhythm vs. Atrial Fibrillation): PahNet achieved an accuracy of 0.9060. While slightly lower than some non-prototype models like CNN (0.9280) and Bi-LSTMAttn (0.9324), it significantly narrowed the performance gap, validating its design.
  • Triple Classification (Normal, AF, Other Rhythms): In this more complex scenario, PahNet achieved a commendable accuracy of 0.8414, with macro precision, recall, and F1-score metrics of 0.8449, 0.8224, and 0.8235, respectively. Crucially, it significantly outperformed other state-of-the-art prototype-based models in this task.
  • Human-Machine Collaboration Impact: The paper shows that refining prototypes with human expertise further improves model performance, with the collaborative PahNet achieving an accuracy of 0.8414 in the three-class task, outperforming its non-collaborative version and other prototype baselines.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Single Heartbeat Focus: The current model analyzes single heartbeats. Future work will incorporate sequential heartbeat analysis to leverage features like RR interval predictability, which are important for clinical diagnosis.
  • Scalability of Collaboration: The human-machine collaboration mechanism needs further optimization for scalability, especially for larger datasets. This may involve exploring more sophisticated machine learning techniques and distributed computing.
  • Expert Quantity Impact: The study used a limited number of experts. Future research will investigate how the quantity of domain experts influences model performance.
  • Outlier and Noise Handling: The model could benefit from more robust outlier and noise mitigation strategies, and the authors plan to integrate advanced filtering techniques.

Summary updated

Abstract

Deep neural networks must address the dual challenge of delivering high-accuracy predictions and providing user-friendly explanations. While deep models are widely used in the field of time series modeling, deciphering the core principles that govern the models’ outputs remains a significant challenge. This is crucial for fostering the development of trusted models and facilitating domain expert validation, thereby empowering users and domain experts to utilize them confidently in high-risk decision-making contexts (e.g., decision-support systems in healthcare). In this work, we put forward a deep prototype learning model that supports interpretable and manipulable modeling and classification of medical time series (i.e., ECG signal). Specifically, we first optimize the representation of single heartbeat data by employing a bidirectional long short-term memory and attention mechanism, and then construct prototypes during the training phase. The final classification outcomes (i.e., normal sinus rhythm, atrial fibrillation, and other rhythm) are determined by comparing the input with the obtained prototypes. Moreover, the proposed model presents a human–machine collaboration mechanism, allowing domain experts to refine the prototypes by integrating their expertise to further enhance the model’s performance (contrary to the human-in-the-loop paradigm, where humans primarily act as supervisors or correctors, intervening when required, our approach focuses on a human–machine collaboration, wherein both parties engage as partners, enabling more fluid and integrated interactions). The experimental outcomes presented herein delineate that, within the realm of binary classification tasks—specifically distinguishing between normal sinus rhythm and atrial fibrillation—our proposed model, albeit registering marginally lower performance in comparison to certain established baseline models such as Convolutional Neural Networks (CNNs) and bidirectional long short-term memory with attention mechanisms (Bi-LSTMAttns), evidently surpasses other contemporary state-of-the-art prototype baseline models. Moreover, it demonstrates significantly enhanced performance relative to these prototype baseline models in the context of triple classification tasks, which encompass normal sinus rhythm, atrial fibrillation, and other rhythm classifications. The proposed model manifests a commendable prediction accuracy of 0.8414, coupled with macro precision, recall, and F1-score metrics of 0.8449, 0.8224, and 0.8235, respectively, achieving both high classification accuracy as well as good interpretability.
BibTeX
@inproceedings{xie2024prototype,
  author = {Jia Xie and Zhu Wang and Zhiwen Yu and Yasan Ding and Bin Guo},
  title = {{Prototype Learning for Medical Time Series Classification via Human-Machine Collaboration}},
  booktitle = {Sensors},
  year = {2024},
  url = {https://openreview.net/forum?id=PkqwvlRVfl},
}
Figure