Neural network ensemble model for prediction of erythrocyte sedimentation rate (ESR) using partial least squares regression

Jaejin Lee, Hyeonji Hong, Jae Min Song, Eunseop YeomScientific 2022 · DOI

TL;DR

Standard erythrocyte sedimentation rate (ESR) tests—a staple for detecting inflammation—require a full hour to complete, creating a bottleneck in clinical workflows. This paper introduces a stacking ensemble deep learning model (GRU-PLSR) that can accurately predict the 60-minute ESR value using only the first 15 minutes of sedimentation data, potentially cutting diagnostic wait times by 75%.

What Problem Does This Paper Address?

The Erythrocyte Sedimentation Rate (ESR) is a non-specific but essential blood test used to monitor inflammatory conditions, ranging from infections to autoimmune diseases. The gold standard, the Westergren method, measures how far red blood cells fall in a vertical tube over exactly one hour.

While simple and inexpensive, this 60-minute window is an obstacle for prompt clinical evaluation. Furthermore, the standard test only provides a single data point at the end, ignoring the complex temporal dynamics of how blood settles. Previous attempts to automate or shorten the test have often suffered from high variability. The authors sought to leverage deep learning to capture the non-linear "tendency" of sedimentation early on to predict the final result.

Methods

The researchers developed a computer vision and machine learning pipeline to tackle the problem:

Overview of the ESR measurement system and data processing workflow, including (a) the experimental setup, (b) sedimentation over time, and (c) the image processing pipeline to extract the sedimentation curve.
  1. Data Acquisition: They used a digital camera to record blood samples in syringes at 0.5-minute intervals for a full hour. Image processing (Otsu’s method) was used to track the interface between plasma and sedimented red cells.
  2. Base Models: They compared three types of Deep Neural Networks (DNNs): Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Because ESR is time-series data, the recurrent architectures (LSTM/GRU) were expected to outperform the standard MLP.
  3. Stacking Ensemble: To improve generalization and prevent overfitting on their relatively small dataset, they used a "stacking" approach. Multiple base models were trained on bootstrapped samples of the data.
  4. Meta-Models: The outputs of these base models were combined using a "meta-model." The authors compared four schemes:
  5. Simple Mean and Median.
  6. LASSO (Least Absolute Shrinkage and Selection Operator).
  7. PLSR (Partial Least Squares Regression).

The inclusion of PLSR is a key technical choice, as it is specifically designed to handle the high correlation (multicollinearity) typically found between the outputs of ensemble base models.

Datasets

  • Primary Dataset: 304 "normal" subjects from the Pusan National University Dental Hospital. This data was split into training (194), validation (50), and test (60) sets.
  • Clinical Validation Set: 6 patients with periodontal disease. This was used to test if a model trained on "normal" sedimentation patterns could generalize to a specific inflammatory disease state.

Key Insights and Ideas

  • The 15-Minute Threshold: The authors found that prediction error (MAPE and RMSE) dropped significantly as the input sequence increased but hit a point of diminishing returns at 15 minutes. This suggests that the "sedimentation phase"—where cells fall most rapidly—is sufficiently captured within the first quarter-hour to forecast the final "packing phase."
  • Sequential Learning Matters: MLP performed poorly because it treats inputs as independent points. ESR is a physical process with three distinct phases (aggregation, sedimentation, and packing); LSTM and GRU succeeded because they "remember" the progression between these phases.
  • PLSR as a Meta-Model: In many ensemble methods, base models are highly correlated because they are solving the same task. The authors demonstrated that PLSR is superior to simple averaging because it projects these correlated variables into a new space of latent variables, effectively filtering out the redundancy.

Key Results

  • Best Performing Model: The GRU–PLSR ensemble was the clear winner. Statistical analysis (Friedman and Nemenyi tests) confirmed it had the lowest mean rank for error across all tested configurations.
  • Accuracy: The GRU-PLSR model achieved a Mean Absolute Percentage Error (MAPE) of approximately 8% using only 15 minutes of data.
  • Generalization: When applied to the periodontitis patients, the model's predictions showed "acceptable agreement" with actual 60-minute measurements via Bland–Altman plots. This indicates the model captures the underlying physics of sedimentation, not just the specific statistics of the training group.

Limitations and Future Work

  • Small Disease Sample: While the model worked on periodontitis patients, the sample size (n=6) is very small. Broader clinical trials with diverse inflammatory conditions (e.g., rheumatoid arthritis, systemic infections) are needed.
  • Feature Engineering: The current model relies solely on the sedimentation curve. The authors note that factors like age, sex, and hematocrit levels also influence ESR. Incorporating these as auxiliary inputs could further refine accuracy.
  • Hardware Integration: The study used a standalone digital camera and syringe setup. To be useful in a clinic, this would need to be integrated into a compact, "point-of-care" device.

Summary updated

Abstract

The erythrocyte sedimentation rate (ESR) is a non-specific blood test for determining inflammatory conditions. However, the long measurement time (60 min) to obtain ESR is an obstacle for a prompt evaluation. In this study, to reduce the measurement time of ESR, deep neural networks (DNNs) were applied to the sedimentation tendency of blood samples. DNNs using multilayer perceptron (MLP), long short-term memory (LSTM), and gated recurrent unit (GRU) were assessed and compared to determine a suitable length of time for the input sequence. To avoid overfitting, a stacking ensemble learning was adopted, which combines multiple models by using a meta model. Four meta models were compared: mean, median, least absolute shrinkage and selection operator, and partial least squares regression (PLSR) schemes. From the empirical results, LSTM and GRU models have better prediction than MLP over sequence lengths of 5 to 20 min. The decrease in [Formula: see text] and [Formula: see text] of GRU and LSTM was attenuated after a sequence length of 15 min, so the input sequence length is determined as 15 min. In terms of the meta model, the statistical comparison suggests that GRU combined with PLSR (GRU-PLSR) is the best case. Then, the GRU-PLSR was tested for prediction of ESR data obtained from periodontitis patients to check its applicability to a specific disease. The Bland-Altman plot shows acceptable agreement between measured and predicted ESR values. Based on the results, the GRU-PLSR can predict ESR with improved performance within 15 min and has potential applicability to ESR data with inflammatory and non-inflammatory conditions.
BibTeX
@inproceedings{lee2022neural,
  author = {Jaejin Lee and Hyeonji Hong and Jae Min Song and Eunseop Yeom},
  title = {{Neural network ensemble model for prediction of erythrocyte sedimentation rate (ESR) using partial least squares regression}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-23174-0},
}
Figure