Systems Engineering-assisted Machine Learning for Biomedical Applications

Abstract

In recent years, advancements in biomedical technology have led to the accumulation of vast amounts of healthcare data. Machine learning (ML) algorithms are now employed to extract valuable insights from data. However, the challenge lies in extracting meaningful information from big data due to irrelevant data and noise. Integrating domain knowledge into ML techniques is crucial to address these challenges and deliver comprehensive and interpretable results.

Speech disorders in children pose diagnostic challenges due to intra- and inter-rater variabilities in auditory perceptual analysis (APA) and manual transcription methods. To overcome these limitations, we explore the utilization of Landmark (LM) analysis with novel knowledge-based features for automatic speech disorder detection. Our systematic study shows nearly a 20% improvement in accuracy, highlighting the effectiveness of these features in classifying speech disorder patients.

A robust framework is proposed that integrates ML techniques with domain knowledge for detecting autism spectrum disorder (ASD) using serum biomarkers. Despite challenges in identifying reliable biomarkers due to protein level variations, our framework outperforms previous methods by integrating feature engineering and selection with linear ML algorithms, achieving high performance in ASD detection. The proposed framework improves the area under the curve (AUC) by 10%, demonstrating its effectiveness in reducing within-class variations.

Furthermore, we investigate the early detection of pulmonary arterial hypertension (PAH) in systemic sclerosis (SSc) patients using proteomic data. Many ML-assisted detection frameworks are limited to the dataset used for training, and they may not perform effectively when applied to different diseases or disorders. This case study underscores the ASD detection framework's effectiveness for detecting various diseases and disorders. The proposed framework achieves over 16% enhancement in PAH detection accuracy from previous detection models.

Our study highlights the efficacy of combining ML with domain knowledge for disorder detection. The feature engineering and selection techniques enhance the robustness and reliability of early detection of disorders, emphasizing the importance of knowledge-guided models for interpretable results.