Learning fine-grained estimation of physiological states from coarse-grained labels by distribution restoration

Zengyi Qin, Jiansheng Chen, Zhenyu Jiang, Xumin Yu, Chunhua Hu, Yu Ma, Suhua Miao, Rongsong ZhouScientific 2020 · DOI

Abstract

Due to its importance in clinical science, the estimation of physiological states (e.g., the severity of pathological tremor) has aroused growing interest in the machine learning community. While the physiological state is a continuous variable, its continuity is lost when the physiological state is quantized into a few discrete classes during recording and labeling. The discreteness introduces misalignment between the true value and its label, meaning that these labels are unfortunately imprecise and coarse-grained. Most previous work did not consider the inaccuracy and directly utilized the coarse labels to train the machine learning algorithms, whose predictions are also coarse-grained. In this work, we propose to learn a precise, fine-grained estimation of physiological states using these coarse-grained ground truths. Established on mathematical rigorous proof, we utilize imprecise labels to restore the probabilistic distribution of precise labels in an approximate order-preserving fashion, then the deep neural network learns from this distribution and offers fine-grained estimation. We demonstrate the effectiveness of our approach in assessing the pathological tremor in Parkinson's Disease and estimating the systolic blood pressure from bioelectrical signals.

Introduction

In clinical settings, experts often label data using discrete scales—for example, rating a Parkinson’s tremor from "slight" to "severe." While these labels are easy to produce, they are "coarse-grained" approximations of a continuous biological reality. Standard machine learning approaches treat this as a classification problem, which creates artificial boundaries in the feature space and fails to capture subtle changes. Conversely, standard regression requires "fine-grained" ground truth, which is often impossible for human annotators to provide accurately. This paper addresses the challenge: How can a model learn to predict continuous, fine-grained values when it has only ever seen discrete, coarse labels?

Methodology

The authors propose a specialized CNN architecture called BioeNet and a training pipeline based on two main pillars: distribution restoration and ordinal scale learning.

Conceptual overview of the proposed method: (a) estimation pipeline, (b-c) comparison of coarse vs. fine-grained estimation, (d) feature space continuity, and (e-f) the transition from discrete to continuous label distributions.

Distribution Restoration

Instead of training the model to match discrete labels, the authors use cubic interpolation on the label histogram to "restore" a smooth, continuous probability density function (PDF). They then use Maximum Mean Discrepancy (MMD) and a kurtosis loss to force the model’s output distribution to match this smooth PDF. This prevents the model from "clumping" its predictions around integer values.

Ordinal Scale Learning (Slack L1 Loss)

To ensure the model respects the natural order of the states (e.g., Level 3 is always more severe than Level 2), they introduce a "Slack L1 loss." This loss function only penalizes the model if its prediction falls outside a certain tolerance range around the coarse label. This provides the "slack" necessary for the model to output a floating-point number (like 2.3) without being forced back to the integer label (2.0).

Experimental Evaluation

The researchers evaluated their approach on three distinct datasets:

  • PD-sEMG: 10,000 sequences of surface electromyography (sEMG) from 147 individuals, labeled with 5 levels of Parkinson’s tremor severity.
  • Synthetic sEMG: A dataset of 100,000 generated signals where the "fine-grained" ground truth was known, allowing for direct validation.
  • Cuff-less Blood Pressure: Bioelectrical signals (PPG and ECG) used to estimate Systolic Blood Pressure (SBP).

Key Results

  • Superior Classification: When rounded to the nearest integer for comparison, BioeNet significantly outperformed state-of-the-art models on the Parkinson’s tremor dataset.
  • Blood Pressure Accuracy: In the SBP task, 92.97% of the model's prediction errors were less than 5 mmHg, outperforming standard ordinal regression techniques (76.56%).
  • Unseen State Prediction: In a "leave-one-out" experiment, the model accurately predicted samples belonging to a middle boundary it had never seen during training, proving it learned the underlying continuous scale.

Discussion and Conclusion

This framework demonstrates that leveraging the ordinal nature of physiological states is key to "hallucinating" missing fine-grained information. By assuming biological transitions are smooth, simple cubic interpolation can create a much richer supervisory signal than raw labels. While the method currently assumes a linear progression, future work could address state bifurcations or apply this to other bio-signals like EEG for high-resolution sleep stage monitoring.

Summary updated

Abstract

Due to its importance in clinical science, the estimation of physiological states (e.g., the severity of pathological tremor) has aroused growing interest in machine learning community. While the physiological state is a continuous variable, its continuity is lost when the physiological state is quantized into a few discrete classes during recording and labeling. The discreteness introduces misalignment between the true value and its label, meaning that these labels are unfortunately imprecise and coarse-grained. Most previous work did not consider the inaccuracy and directly utilized the coarse labels to train the machine learning algorithms, whose predictions are also coarse-grained. In this work, we propose to learn a precise, fine-grained estimation of physiological states using these coarse-grained ground truths. Established on mathematical rigorous proof, we utilize imprecise labels to restore the probabilistic distribution of precise labels in an approximate order-preserving fashion, then the deep neural network learns from this distribution and offers fine-grained estimation. We demonstrate the effectiveness of our approach in assessing the pathological tremor in Parkinson's Disease and estimating the systolic blood pressure from bioelectrical signals.
BibTeX
@inproceedings{qin2020learning,
  author = {Zengyi Qin and Jiansheng Chen and Zhenyu Jiang and Xumin Yu and Chunhua Hu and Yu Ma and Suhua Miao and Rongsong Zhou},
  title = {{Learning fine-grained estimation of physiological states from coarse-grained labels by distribution restoration}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-79007-5},
}

Loading related content...

Figure