Tri-MTL: A Triple Multitask Learning Approach for Respiratory Disease Diagnosis

TL;DR

Tri-MTL addresses the challenge of diagnosing respiratory diseases by jointly modeling lung sounds, disease labels, and patient metadata using a pretrained Audio Spectrogram Transformer. The approach extends traditional multitask learning with a metadata classification task, enabling richer representations and improved performance on both lung-sound classification and disease diagnosis on the ICBHI dataset. Key contributions include a simple MTL framework with AST, the Tri-MTL extension incorporating metadata attributes, and comprehensive results showing that metadata, especially stethoscope information, boosts performance. The findings suggest practical value for clinical decision support by enhancing diagnostic accuracy and efficiency in respiratory care.

Abstract

Auscultation remains a cornerstone of clinical practice, essential for both initial evaluation and continuous monitoring. Clinicians listen to the lung sounds and make a diagnosis by combining the patient's medical history and test results. Given this strong association, multitask learning (MTL) can offer a compelling framework to simultaneously model these relationships, integrating respiratory sound patterns with disease manifestations. While MTL has shown considerable promise in medical applications, a significant research gap remains in understanding the complex interplay between respiratory sounds, disease manifestations, and patient metadata attributes. This study investigates how integrating MTL with cutting-edge deep learning architectures can enhance both respiratory sound classification and disease diagnosis. Specifically, we extend recent findings regarding the beneficial impact of metadata on respiratory sound classification by evaluating its effectiveness within an MTL framework. Our comprehensive experiments reveal significant improvements in both lung sound classification and diagnostic performance when the stethoscope information is incorporated into the MTL architecture.

Figures (2)

• Figure 1: Illustration of the proposed MTL approaches for joint lung sound classification and disease diagnosis. $(a)$ Hard parameter sharing utilizes a single shared audio feature encoder to extract representations for both tasks, followed by separate classifiers for lung sound classification and disease diagnosis. $(b)$ Soft parameter sharing introduces two task-specific encoders, one for lung sound classification and the other for disease diagnosis, with a regularization loss encouraging shared representation learning between them. $(c)$ Tri-MTL extends both lung sound classification and disease diagnosis tasks by incorporating an additional metadata classification task, enhancing the model’s ability to leverage metadata attributes. For the Tri-MTL, both hard and soft parameter sharing strategies are considered.
• Figure 2: t-SNE visualizations illustrating how models trained in isolation for a single task organize their feature spaces, highlighting why MTL benefits disease diagnosis more than lung sound classification.