TL;DR
This paper introduces a novel approach to detecting delirium in elderly patients by combining Bispectral EEG (BSEEG) with Topological Data Analysis (TDA). The TDA enhancement significantly improves the accuracy of delirium detection compared to BSEEG alone, offering a more objective and practical tool for clinical settings.
What Problem Does This Paper Address?
Delirium is a common, costly, and dangerous condition in elderly inpatients, yet it remains significantly underdiagnosed and undertreated. Current diagnostic methods, like the Confusion Assessment Method (CAM) and Delirium Rating Scale (DRS), are often subjective, time-consuming, and challenging to implement in busy hospital environments, leading to suboptimal sensitivity. While Electroencephalography (EEG) shows promise for objective detection, traditional EEG systems are cumbersome, expensive, and require expert interpretation, hindering widespread clinical use. The authors previously developed a simplified, portable BSEEG algorithm for delirium detection but sought to further improve its accuracy, particularly for challenging cases.
Methods
The study builds upon a previously developed Bispectral EEG (BSEEG) algorithm. BSEEG calculates a score based on the ratio of low-frequency to high-frequency power in EEG signals, aiming to capture the "diffuse slowing" characteristic of delirium. This study introduces Topological Data Analysis (TDA) as a complementary method. TDA, specifically persistent homology, is applied to the EEG data to extract global features and topological characteristics of the signal's "shape." The TDA score is derived from a 1-dimensional Betti curve, reflecting the irregularity of the time series.
The process involved:
- Signal Pre-processing: Raw EEG signals were filtered to remove noise (e.g., electromagnetic interference) and artifacts (e.g., EMG), and band-pass filtered to isolate relevant EEG frequencies (0.5-20 Hz). Only signals with at least 30 seconds of usable data were included.
- BSEEG Score Calculation: This involved a power spectral density (PSD) ratio between 3 Hz and 10 Hz, as described in prior work.
- TDA Score Calculation: Persistent homology was applied to time-delay embeddings of the EEG data. The score was derived from the Betti curve.
- Comparison: BSEEG and TDA scores were compared between delirious and non-delirious patient groups.
Datasets
The study utilized two independent cohorts of elderly inpatients:
- 1st Cohort: 274 subjects enrolled at the University of Iowa Hospitals and Clinics (UIHC) between January 30, 2016, and October 30, 2017. Of these, 102 were categorized as delirious.
- 2nd Cohort: 265 subjects recruited at UIHC between January 16, 2019, and August 19, 2019. Of these, 42 were categorized as delirious.
The study used two different portable EEG devices for data collection across the cohorts.
Key Insights and Ideas
- TDA Captures Subtle Delirium Patterns: The most striking insight is TDA's ability to identify delirium in cases where BSEEG alone might miss it. This is illustrated by a case where EEG signals were not obviously slow to the human eye, resulting in a low BSEEG score, but TDA successfully identified it as a delirium case with a high score. This suggests TDA can capture more nuanced or complex alterations in brain activity associated with delirium.
- Improved Objectivity and Practicality: By enhancing an already portable and simplified EEG method (BSEEG), the combined TDA-BSEEG approach offers a more objective and potentially more accurate tool for delirium detection. This addresses the limitations of subjective clinical assessments and the impracticality of traditional EEG.
- Robustness Across Cohorts and Devices: The consistent improvement observed with TDA across two independent cohorts, using different EEG devices, lends significant credibility to the findings and suggests the approach is robust and generalizable.
Key Results
The study demonstrated a significant improvement in delirium detection performance with the addition of TDA:
-
Area Under the Curve (AUC):
- 1st Cohort (Fp1-A1): BSEEG AUC = 0.72, TDA AUC = 0.82 (p=0.006)
- 1st Cohort (Fp2-A2): BSEEG AUC = 0.76, TDA AUC = 0.84 (p=0.015)
- 2nd Cohort (Fp1-A1): BSEEG AUC = 0.67, TDA AUC = 0.78 (p=0.014) TDA consistently yielded significantly higher AUCs than BSEEG alone.
-
Specificity Improvement: When sensitivity was fixed at 0.80:
- 1st Cohort (Fp1-A1): BSEEG specificity = 0.48, TDA specificity = 0.66
- 1st Cohort (Fp2-A2): BSEEG specificity = 0.54, TDA specificity = 0.72
- 2nd Cohort (Fp1-A1): BSEEG specificity = 0.46, TDA specificity = 0.62 TDA significantly improved specificity, ranging from a 16% to 26% increase compared to BSEEG.
Limitations and Future Work
- Artifact Contamination: Raw EEG data is prone to artifacts (e.g., EMG) that can overlap with EEG frequency bands, making complete removal challenging with current filtering techniques. The authors suggest improvements in EEG hardware and algorithmic approaches are needed.
- Sample Size Reduction: The rigorous pre-processing required for TDA analysis led to a reduction in the number of usable samples (60-80% of original data), which is not ideal for clinical practice. Future work will focus on finding a balance between signal quality and sample inclusion.
- Parameter Optimization: TDA analysis involves several adjustable parameters. While the study found consistent benefits with the chosen parameters across datasets, further validation on independent samples is needed to confirm these parameters are universally optimal.
- Computational Intensity: TDA is currently computationally intensive, requiring analysis on a PC rather than real-time bedside implementation. The authors are actively working on optimizing the algorithm's speed for potential handheld device integration.
- Generalizability: The study was conducted at a single institution in the US, with participants primarily being non-Hispanic White. Future research should aim to validate these findings in more diverse populations and institutions.
Summary updated