Toward Automated Classification of Pathological Transcranial Doppler Waveform Morphology via Spectral Clustering

Samuel Thorpe, Corey M. Thibeault, Nicolas Canac, Kian Jalaleddini, Amber Y. Dorn, Seth J. Wilk, Thomas Devlin, Fabien Scalzo, Robert Hamilton

Abstract

Cerebral Blood Flow Velocity (CBFV) waveforms acquired via Transcranial Doppler (TCD) can provide evidence for cerebrovascular occlusion and stenosis. Thrombolysis in Brain Ischemia (TIBI) flow grades are widely used for this purpose, but require subjective assessment by expert evaluators to be reliable. In this work we seek to determine whether TCD morphology can be objectively assessed using an unsupervised machine learning approach to waveform categorization. TCD beat waveforms were recorded at multiple depths from the Middle Cerebral Arteries of 106 subjects; 33 with CTA-confirmed Large Vessel Occlusion (LVO). From each waveform, three morphological variables were extracted, quantifying absolute peak onset, number/prominence of auxiliary peaks, and systolic canopy length. Spectral clustering identified groups implicit in the resultant three-dimensional feature space, with gap-statistic criteria establishing the optimal cluster number. We found that gap-statistic disparity was maximized at four clusters, referred to as flow types I, II, III, and IV. Types I and II were primarily composed of control subject waveforms, whereas types III and IV derived mainly from LVO patients. Cluster morphologies for types I and IV aligned clearly with Normal and Blunted TIBI flows, respectively. Types II and III represented commonly observed flow-types not delineated by TIBI, which nonetheless deviate quantifiably from normal and blunted flows. We conclude that important morphological variability exists beyond that currently quantified by TIBI in populations experiencing or at-risk for acute ischemic stroke, and posit that the observed flow-types provide the foundation for objective methods of real-time automated flow type classification.

Introduction

Transcranial Doppler (TCD) ultrasound is a non-invasive diagnostic modality used to measure Cerebral Blood Flow Velocity (CBFV). It is particularly valuable in emergency neurology for detecting Large Vessel Occlusions (LVO) in patients suspected of having an acute ischemic stroke. Currently, the clinical standard for interpreting these waveforms is the Thrombolysis in Brain Ischemia (TIBI) scale. While effective, the TIBI scale relies on the subjective qualitative assessment of waveform shapes by expert clinicians. This dependency limits the utility of TCD in pre-hospital settings or in facilities without specialized neurovascular staff.

This research investigates whether unsupervised machine learning can overcome these limitations by establishing an objective, data-driven taxonomy of CBFV waveforms. By identifying natural groupings within the data, the study aims to provide a framework for automated, real-time diagnostic tools.

Methodology

Data Collection and Preprocessing

The study analyzed 996 average beat waveforms collected from 106 subjects. This cohort included 33 patients with CTA-confirmed LVO, 33 in-hospital controls (IHC), and 40 out-of-hospital healthy controls (OHC). To focus exclusively on the morphology (shape) of the pulses, all waveforms were normalized for both time and velocity, removing the confounding effects of heart rate and absolute flow speed.

Feature Extraction

Three key morphological features were extracted from each normalized waveform to represent its geometric properties:

  1. Onset: The time required to reach maximum systolic velocity.
  2. Peaks: A weighted count of primary and secondary auxiliary peaks, reflecting the complexity of the systolic phase.
  3. Canopy: The duration of the systolic peak, or the "width" of the wave's crest.
Figure 1: Extraction of morphological features (onset, peaks, and canopy) from example waveforms across different subject groups.

Clustering and Optimization

The researchers utilized spectral clustering to group the waveforms within a three-dimensional feature space. To determine the optimal number of clusters ($k$), they employed the gap statistic. This method compares the total intra-cluster dispersion of the data against a null distribution of random data, identifying the point where the gap between the two is maximized.

Results

The gap statistic analysis revealed that the data naturally organized into four distinct clusters, designated as Flow Types I, II, III, and IV. These clusters showed significant clinical relevance:

  • Type I (Normal): This cluster was the most frequent (400 waveforms) and was dominated by control subjects (82%). Its morphology, characterized by rapid onset and distinct peaks, aligns with the "Normal" TIBI grade.
  • Type II (Late Onset/Normal Peaks): Also primarily composed of controls (80%), this type exhibited a slightly delayed systolic peak compared to Type I.
  • Type III (Early Onset/Weak Peaks): This cluster was highly pathological; 95% of the waveforms in this group were from LVO patients. This represents a critical finding, as Type III morphology is not explicitly defined in the traditional TIBI scale but is a strong indicator of occlusion.
  • Type IV (Blunted): Associated with LVO (60%), these waveforms showed the classic "blunted" appearance—delayed acceleration and a lack of distinct auxiliary peaks.
Figure 2: Average morphologies for the four discovered clusters and their corresponding clinical distributions (Control vs. LVO).

Discussion

The identification of four clusters suggests that the traditional six-grade TIBI scale may not fully capture the natural variance in blood flow morphology. Specifically, the emergence of Type III as a high-specificity marker for LVO highlights the power of unsupervised learning to discover "quantifiably deviant" patterns that human observers might overlook.

While the study successfully categorized waveforms based on shape, the researchers noted that velocity normalization removed information about absolute speed, which is a component of some TIBI grades (e.g., Grade 3, "Dampened"). Future iterations of this pipeline could integrate absolute velocity metrics or the Velocity Curvature Index (VCI) to enhance diagnostic robustness.

Conclusion

This work demonstrates that spectral clustering of morphological features provides an objective and clinically relevant method for classifying TCD waveforms. By moving beyond subjective scales like TIBI, this approach paves the way for automated diagnostic systems that can provide rapid, reliable stroke triage in both hospital and pre-hospital environments.

Summary updated

Abstract

Abstract Cerebral Blood Flow Velocity waveforms acquired via Transcranial Doppler (TCD) can provide evidence for cerebrovascular occlusion and stenosis. Thrombolysis in Brain Ischemia (TIBI) flow grades are widely used for this purpose, but require subjective assessment by expert evaluators to be reliable. In this work we seek to determine whether TCD morphology can be objectively assessed using an unsupervised machine learning approach to waveform categorization. TCD beat waveforms were recorded at multiple depths from the Middle Cerebral Arteries of 106 subjects; 33 with CTA-confirmed Large Vessel Occlusion (LVO). From each waveform, three morphological variables were extracted, quantifying absolute peak onset, number/prominence of auxiliary peaks, and systolic canopy length. Spectral clustering identified groups implicit in the resultant three-dimensional feature space, with gap-statistic criteria establishing the optimal cluster number. We found that gap-statistic disparity was maximized at four clusters, referred to as flow types I, II, III, and IV. Types I and II were primarily composed of control subject waveforms, whereas types III and IV derived mainly from LVO patients. Cluster morphologies for types I and IV aligned clearly with Normal and Blunted TIBI flows, respectively. Types II and III represented commonly observed flow-types not delineated by TIBI, which nonetheless deviate quantifiably from normal and blunted flows. We conclude that important morphological variability exists beyond that currently quantified by TIBI in populations experiencing or at-risk for acute ischemic stroke, and posit that the observed flow-types provide the foundation for objective methods of real-time automated flow type classification.
BibTeX
@inproceedings{thorpe2020toward,
  author = {Samuel Thorpe and Corey M. Thibeault and Nicolas Canac and Kian Jalaleddini and Amber Y. Dorn and Seth J. Wilk and Thomas Devlin and Fabien Scalzo and Robert Hamilton},
  title = {{Toward Automated Classification of Pathological Transcranial Doppler Waveform Morphology via Spectral Clustering}},
  booktitle = {PLoS},
  year = {2020},
  doi = {10.1371/journal.pone.0228642},
}
Figure