[Paper Review] Towards Wide Learning: Experiments in Healthcare

Snehasis Banerjee, Tanushyam Chattopadhyay+5 more|arXiv (Cornell University)|Dec 17, 2016

Phonocardiography and Auscultation Techniques5 references3 citations

TL;DR

This paper proposes a Wide Learning architecture to automate feature engineering in healthcare machine learning, using hierarchical feature listing and hybrid feature selection (mRMR and MRMS) with SVM. It achieves 94.38% of the PhysioNet Challenge winner’s accuracy in just 3 person-days, outperforming state-of-the-art methods on two datasets while delivering interpretable, human-understandable features with drastically reduced effort.

ABSTRACT

In this paper, a Wide Learning architecture is proposed that attempts to automate the feature engineering portion of the machine learning (ML) pipeline. Feature engineering is widely considered as the most time consuming and expert knowledge demanding portion of any ML task. The proposed feature recommendation approach is tested on 3 healthcare datasets: a) PhysioNet Challenge 2016 dataset of phonocardiogram (PCG) signals, b) MIMIC II blood pressure classification dataset of photoplethysmogram (PPG) signals and c) an emotion classification dataset of PPG signals. While the proposed method beats the state of the art techniques for 2nd and 3rd dataset, it reaches 94.38% of the accuracy level of the winner of PhysioNet Challenge 2016. In all cases, the effort to reach a satisfactory performance was drastically less (a few days) than manual feature engineering.

Motivation & Objective

• To automate the time-consuming and expertise-intensive process of feature engineering in healthcare machine learning.
• To develop a hierarchical, interpretable feature recommendation system that supports causality analysis and domain knowledge integration.
• To reduce the effort required to build high-performing ML models in clinical analytics from months to days.
• To outperform state-of-the-art methods, including deep learning and PCA, on real-world healthcare datasets with interpretable features.

Proposed method

• The method uses a hierarchical feature listing: Level 0 includes time-domain, Fourier-based (STFT), and wavelet-based (DWT) features; Level 1 adds spectral, statistical, and peak-trough features; Level 2 computes ratios and derivatives of Level 1 features.
• Automated mother wavelet selection is performed using the energy-to-entropy ratio to optimize DWT feature extraction.
• Feature selection combines mRMR (minimizing redundancy, maximizing relevance via mutual information) and MRMS (fuzzy-rough set-based relevance and significance scoring) to identify optimal feature subsets.
• The final feature set is the union of mRMR and MRMS outputs, with size k selected iteratively and evaluated using SVM with various kernels.
• Performance is evaluated using accuracy on hidden test sets after p-fold cross-validation (p=5–10), with Theano-based MLP and dropout used for comparison.
• The system supports extensibility by allowing plug-in of new feature extraction algorithms, enabling a collaborative ecosystem for feature engineering.

Experimental results

Research questions

• RQ1 Can an automated feature recommendation system achieve performance comparable to manual feature engineering in healthcare ML tasks?
• RQ2 Does the proposed Wide Learning architecture with hierarchical feature engineering outperform state-of-the-art methods, including deep learning and PCA, on limited healthcare datasets?
• RQ3 Can the method significantly reduce the time and expertise required for model development while maintaining or improving accuracy?
• RQ4 To what extent are the recommended features interpretable and suitable for clinical causality analysis?
• RQ5 How does the integration of mRMR and MRMS improve feature selection compared to individual techniques?

Key findings

• The proposed method achieved 84% accuracy on the PhysioNet Challenge 2016 dataset, reaching 94.38% of the winner’s accuracy (0.89) with only 3 person-days of effort, compared to 120 person-days for manual feature engineering.
• On the MIMIC-II blood pressure classification dataset, the method achieved 87.8% accuracy, surpassing the state-of-the-art (79.5%) and outperforming both MLP (50%) and PCA+SVM (62.5%) approaches.
• For the in-house emotion classification dataset, the method achieved 90.9% accuracy, exceeding the state-of-the-art (82.3%) and significantly outperforming MLP (50%) and PCA+SVM (50%) methods.
• The method reduced development effort to just 2–3 person-days per dataset, compared to 90–130 person-days for manual feature engineering, demonstrating a substantial efficiency gain.
• Unlike deep learning and PCA, the proposed method generates interpretable features, enabling clinical insight and causality analysis, which is critical in healthcare applications.
• The hybrid use of mRMR and MRMS for feature selection improved performance by capturing complementary aspects of relevance and significance, outperforming either method alone.