PATTERN: Pain Assessment for paTients who can't TEll using Restricted Boltzmann machiNe

Lei Yang, Shuang Wang, Xiaoqian Jiang, Samuel Cheng, Hyeon-Eui KimBMC Medical Informatics Decis. Mak 2016 · OpenReview

TL;DR

This paper introduces PATTERN, a novel machine learning approach using Restricted Boltzmann Machines (RBMs) to predict pain in patients who cannot verbally communicate their discomfort. By modeling the complex relationships within physiological data, PATTERN demonstrates improved accuracy over traditional methods like PCA and LDA, offering a promising tool for objective pain assessment in critical care settings.

What Problem Does This Paper Address?

Accurately assessing pain is crucial for effective patient care, yet it's a significant challenge for individuals unable to self-report, such as those with severe brain injuries or in minimally responsive states. Current pain management often relies on subjective patient accounts, which are unavailable in these cases. While physiological signals offer a potential objective measure, practical and reliable methods for translating these signals into pain assessments are scarce. This paper aims to bridge this gap by exploring whether patterns within electronic physiological signals can reliably predict a patient's pain state, even without direct verbal feedback.

Methods

The core of the PATTERN approach involves using Restricted Boltzmann Machines (RBMs), a type of deep learning model, to learn the underlying patterns in patient data. The process begins by converting raw, n-dimensional numerical physiological data into binary representations. These binary features, along with the pain state (also represented in binary), are then used to construct an (n+1)-dimensional dataset. An RBM is trained on this data to model the joint probability distribution of all variables.

A key innovation is the use of a discriminant RBM for classification. This allows the model to directly compute the probability of a pain state given the observed physiological data. The authors also describe a pre-processing step where numerical features are converted to binary using a Gaussian probability distribution to account for individual variations in "normal" physiological states. For comparison, they also evaluate Principal Component Analysis (PCA) followed by Support Vector Machines (SVM), and Linear Discriminant Analysis (LDA).

Datasets

The study utilized a dataset collected from 7384 patients, comprising 937,461 flowsheet entries. For their experiments, the researchers selected 78 specific data items that were documented with sufficient frequency and in continuous numerical values. They synchronized this time-series data using MATLAB. For the core experiments, they focused on four individuals who had over 1000 recorded pain labels. The number of available data items for these four patients varied between 22 and 28. Approximately 20% of the data for each patient was randomly selected as the test set.

Key Insights and Ideas

  • Hidden Pain Signals: The research suggests that physiological signals, when analyzed with sophisticated models like RBMs, contain discernible patterns that correlate with pain, even in non-communicative patients.
  • RBMs for Complex Dependencies: RBMs prove effective at modeling the intricate, non-linear dependencies within the binary physiological data, outperforming linear methods that might miss these nuances.
  • Discriminant RBM Advantage: The discriminant RBM approach, which directly leverages the model's probabilistic nature for classification, shows superior performance compared to simply using RBM's hidden representations with a separate classifier.
  • Binary Representation Utility: While converting continuous physiological data to binary might seem like a simplification, the authors found it effective for their RBM model, particularly when interpreted as "normal" vs. "abnormal" states.

Key Results

The study found that the discriminant RBM approach consistently achieved better accuracy across all four tested patients compared to PCA+SVM and LDA. Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values demonstrated this superiority. For instance, in one patient, the RBM achieved an AUC of 0.8813, significantly higher than PCA+SVM (0.76504) and LDA (0.82033). Table 2 further quantifies these results, showing RBM achieving the best sensitivity, specificity, and accuracy in multiple cases.

Limitations and Future Work

The authors acknowledge several limitations:

  • Ignoring Temporal Data: The current approach treats data as non-temporal, which is a significant loss of information. Future work should incorporate time-series analysis.
  • Binary Representation Sufficiency: While effective, a purely binary representation might not capture the full spectrum of physiological changes (e.g., blood pressure being high or low). The authors plan to explore richer representations, such as using softmax visible units in RBMs.
  • Small Dataset Size: The experiments were conducted on a relatively small dataset for individual patients. Future research will focus on training models with larger datasets and fine-tuning parameters for individual patients to minimize data loss and improve generalization.

Summary updated

Abstract

Accurately assessing pain for those who cannot make self-report of pain, such as minimally responsive or severely brain-injured patients, is challenging. In this paper, we attempted to address this challenge by answering the following questions: (1) if the pain has dependency structures in electronic signals and if so, (2) how to apply this pattern in predicting the state of pain. To this end, we have been investigating and comparing the performance of several machine learning techniques. We first adopted different strategies, in which the collected original n-dimensional numerical data were converted into binary data. Pain states are represented in binary format and bound with above binary features to construct (n + 1) -dimensional data. We then modeled the joint distribution over all variables in this data using the Restricted Boltzmann Machine (RBM). Seventy-eight pain data items were collected. Four individuals with the number of recorded labels larger than 1000 were used in the experiment. Number of avaliable data items for the four patients varied from 22 to 28. Discriminant RBM achieved better accuracy in all four experiments. The experimental results show that RBM models the distribution of our binary pain data well. We showed that discriminant RBM can be used in a classification task, and the initial result is advantageous over other classifiers such as support vector machine (SVM) using PCA representation and the LDA discriminant method.
BibTeX
@inproceedings{yang2016pattern,
  author = {Lei Yang and Shuang Wang and Xiaoqian Jiang and Samuel Cheng and Hyeon-Eui Kim},
  title = {{PATTERN: Pain Assessment for paTients who can't TEll using Restricted Boltzmann machiNe}},
  booktitle = {BMC Medical Informatics Decis. Mak},
  year = {2016},
  url = {https://openreview.net/forum?id=jo1FKLPdiK},
}
Figure