Introduction
In healthcare analytics, the deployment of machine learning models faces two primary obstacles: the "feature engineering bottleneck" and the "black box" problem. Traditional feature engineering requires extensive domain expertise and months of iterative development. Meanwhile, high-performing models like Deep Learning often lack interpretability, making it difficult for clinicians to trust or validate their outputs for patient care. This paper introduces a "Wide Learning" architecture designed to automate feature engineering while maintaining the transparency required for medical applications.
Methodology: The Wide Learning Architecture
The proposed Wide Learning framework shifts away from the abstract layers of Deep Learning, focusing instead on stacking physically meaningful signal processing techniques. The architecture is divided into two primary modules: Feature Listing and Feature Selection.
Feature Listing
Features are extracted hierarchically across three levels:
- Level 0 (Basic Features): Fundamental transforms such as Time Domain (TD), Short-Time Fourier Transform (STFT), and Discrete Wavelet Transform (DWT).
- Level 1 (Dependent Features): Statistical properties, spectral features, and peak/trough detection derived from Level 0.
- Level 2 (Derived Features): Ratios and derivatives of features from the previous levels.
Feature Selection and Recommendation
To manage the exponential growth of potential feature combinations, the system employs an iterative selection process using mRMR (Minimum Redundancy Maximum Relevance) and MRMS (Maximal Relevance Maximum Significance) algorithms. A "simplicity first" rule governs the recommendation: the system attempts to meet performance targets using the lowest possible feature level before escalating to more complex ones. To ensure computational feasibility, a hardware-aware regularizer ('c') limits the complexity of feature combinations.
Experimental Setup
The researchers evaluated the Wide Learning (WIDE) method on three physiological signal datasets:
- PhysioNet Challenge 2016 (D1): 3,153 heart sound recordings (PCG) for normal/abnormal classification.
- MIMIC-II (D2): Blood pressure classification using photoplethysmogram (PPG) signals.
- Emotion Classification (D3): An in-house PPG dataset for classifying "happy" vs. "sad" emotional states.
Results and Discussion
The WIDE method was benchmarked against Multi-Layer Perceptrons (MLP), PCA+SVM, and existing State-of-the-Art (SoA) manual methods.
Performance and Interpretability
WIDE achieved superior accuracy on the MIMIC-II (0.878) and Emotion (0.909) datasets, outperforming both black-box models and manual SoA methods. On the PhysioNet challenge, it remained competitive with an accuracy of 0.84. Crucially, WIDE provides interpretable features—such as heart rate variability or peak-to-trough ratios—allowing for clinical validation, a feature missing in MLP and PCA-based approaches.
Efficiency Gains
The most significant advantage of Wide Learning is the reduction in human effort. For the PhysioNet dataset, manual feature engineering by experts typically required 120 person-days. The automated WIDE approach achieved comparable results in just 3 days, representing a massive leap in development efficiency.
Conclusion
Wide Learning offers a promising path toward automating the machine learning pipeline in healthcare without sacrificing interpretability. By hierarchically extracting and selecting physically meaningful features, the system achieves state-of-the-art performance while drastically reducing the time from raw data to actionable insights. Future work will focus on expanding this approach to other domains, such as machine prognostics, and integrating medical knowledge bases to enhance causality analysis.
Summary updated