HSDreport: Heart Sound Diagnosis with Echocardiography Reports

TL;DR

HSDreport is introduced, a new benchmark for HSD, which mandates the direct utilization of heart sounds obtained from auscultation to predict echocardiography reports, and significantly outperforms traditional HSD approaches and existing multimodal LLMs in detecting key abnormalities in heart sounds.

Abstract

Heart sound auscultation holds significant importance in the diagnosis of congenital heart disease. However, existing methods for Heart Sound Diagnosis (HSD) tasks are predominantly limited to a few fixed categories, framing the HSD task as a rigid classification problem that does not fully align with medical practice and offers only limited information to physicians. Besides, such methods do not utilize echocardiography reports, the gold standard in the diagnosis of related diseases. To tackle this challenge, we introduce HSDreport, a new benchmark for HSD, which mandates the direct utilization of heart sounds obtained from auscultation to predict echocardiography reports. This benchmark aims to merge the convenience of auscultation with the comprehensive nature of echocardiography reports. First, we collect a new dataset for this benchmark, comprising 2,275 heart sound samples along with their corresponding reports. Subsequently, we develop a knowledge-aware query-based transformer to handle this task. The intent is to leverage the capabilities of medically pre-trained models and the internal knowledge of large language models (LLMs) to address the task's inherent complexity and variability, thereby enhancing the robustness and scientific validity of the method. Furthermore, our experimental results indicate that our method significantly outperforms traditional HSD approaches and existing multimodal LLMs in detecting key abnormalities in heart sounds.

Figures (5)

• Figure 1: Comparison between previous works and ours. In previous works, heart sound diagnosis (HSD) was treated as a multi-class problem with two to five categories. Our new benchmark, starting from echocardiography reports, treats HSD as a twelve-category multi-label task. Furthermore, we have developed a knowledge-aware, query-based transformer approach that utilizes medical descriptions to address this novel benchmark.
• Figure 2: Framework overview of HSDreport, which consists of (a) Cardiac Entity Extraction to filter the hot words from the abnormal cardiac entities and verify the existence of them with GPT-4, (b) Training stage to train our model with extracted entity definitions and semantic descriptions paired by the heart sounds, and (c) Inference stage to obtain the diagnosis for a specific entity with various definitions derived from GPT-4.
• Figure 3: Data distribution. The total number of samples is 2275, and the numbers in the figure represent the number of positives in that category.
• Figure 4: ROC curves of four classes in the benchmark: VSD, PDA, Shunt, and Hypertrophy.
• Figure 5: The distribution of F1 scores for VSD, PDA, Shunt, and Hypertrophy categories.