Heterogeneity in the diagnosis and prognosis of ischemic stroke subtypes: 9-year follow-up of 22,000 cases in Chinese adults

Matthew Chun, Haiqiang Qin, Iain Turnbull, Sam Sansome, Simon Gilbert, Alex Hacker, Neil Wright, Tingting Zhu, David A. Clifton, Derrick Bennett, Yu Guo, Pei Pei, Jun Lv, Canqing Yu, Ling Yang, Liming Li, Yan Lu, Zhengming Chen, Benjamin J. Cairns, Yiping Chen, Robert ClarkeInternational Journal of Stroke 2023 · DOI

TL;DR

This study tackles the challenge of accurately classifying ischemic stroke (IS) subtypes, crucial for prognosis and treatment, especially in populations with limited diagnostic resources. By applying machine learning to a large Chinese cohort, they demonstrate its effectiveness in classifying strokes with incomplete investigations, revealing significant prognostic differences between subtypes and highlighting the heterogeneity of IS.

What Problem Does This Paper Address?

Accurately classifying ischemic stroke (IS) into etiological subtypes (like large artery atherosclerosis, cardioaortic embolism, or small artery occlusion) is vital for understanding prognosis and guiding treatment. However, in large population studies or resource-limited settings, obtaining complete diagnostic information (imaging, cardiac tests, vascular investigations) is often impossible, leading to a significant proportion of "undetermined etiology" cases. This lack of classification hinders our ability to accurately assess risk and tailor interventions. Furthermore, the prevalence and prognostic implications of these subtypes can vary significantly across different populations, such as Chinese versus Western populations, necessitating region-specific research.

Methods

The researchers employed a two-pronged approach. First, they used a modified Causative Classification System for Ischemic Stroke (CCS) to classify IS cases with available diagnostic data into "evident," "probable," or "possible" categories. For the substantial number of cases with "undetermined etiology" and incomplete investigations, they developed a machine learning (ML) model. This model utilized a multiclass logistic regression (LR) algorithm, trained on baseline risk factors and screening for cardioaortic sources of embolism from a subset of the data. Feature selection for the LR model was optimized using L1 regularization to maximize the area under the receiver-operating characteristic curve (AUC). The performance of the ML-predicted subtypes was then compared to the etiologically classified subtypes in terms of subsequent stroke risk and all-cause mortality.

Datasets

The study utilized data from the China Kadoorie Biobank (CKB), a prospective study of 512,726 Chinese adults aged 30-79 years, enrolled between 2004-2008, with follow-up data available until January 2018. This dataset included detailed information on health, lifestyle, and physical measurements, along with incident IS cases confirmed by clinical adjudication of medical records. A total of 22,216 incident IS cases were identified. The ML model was trained and tested on a subset of these cases with evident or probable etiology, and then applied to cases with undetermined etiology and incomplete investigations. No new benchmark datasets were introduced; the study leverages existing CKB data.

Key Insights and Ideas

  • ML as a powerful tool for classification in data-scarce scenarios: The study convincingly demonstrates that ML can effectively bridge the gap in classifying IS subtypes when traditional diagnostic methods are incomplete, offering a probabilistic approach to risk stratification.
  • Prognostic heterogeneity is significant: Even within the broad category of ischemic stroke, the subtypes exhibit stark differences in their long-term outcomes, with cardioaortic embolism (CE) posing the highest risk for subsequent stroke and mortality, followed by large artery atherosclerosis (LAA) and small artery occlusion (SAO).
  • Population-specific differences are critical: The distribution of IS subtypes in China (high SAO, low CE) differs markedly from Western populations, underscoring the need for localized research and potentially tailored diagnostic and treatment strategies.
  • ML predictions align with clinical classifications, with a caveat: While ML-predicted subtypes showed comparable stroke risks to clinically classified ones, their all-cause mortality rates were consistently higher. This suggests that the "undetermined etiology" group, even when classified by ML, might represent a higher-risk population, potentially due to differences in access to care or diagnostic investigations in the real-world settings from which the data were drawn.

Key Results

  • Out of 22,216 primary IS cases, 70% had a determined causative subtype, with 48% classified as evident or probable etiology.
  • Among evident and probable IS subtypes, SAO accounted for 66%, LAA for 32%, and CE for 2%.
  • CE had the highest 5-year subsequent stroke (43.5%) and all-cause mortality (40.7%) rates, followed by LAA (43.2% and 17.4%) and SAO (38.1% and 11.1%).
  • The ML model achieved high discrimination for CE (AUC: 0.99) but more modest discrimination for LAA (AUC: 0.67) and SAO (AUC: 0.70) in unseen cases.
  • ML-classified IS subtypes showed comparable subsequent stroke rates to etiologically classified subtypes.
  • ML-classified cases exhibited generally concordant, but slightly higher, all-cause mortality rates compared to etiologically classified IS cases.

Limitations and Future Work

The authors acknowledge several limitations:

  • Representativeness of the CKB: The CKB cohort may not be fully representative of the entire Chinese population, potentially limiting the generalizability of findings to other regions or demographics.
  • Missing baseline data: Information on atrial fibrillation (AF) and blood lipids was not available for all participants.
  • Impact of region on classification: The inclusion of geographic region as a variable in the ML model might limit its applicability to other regions.
  • Underestimation of mortality: Fatal stroke cases reported only to death registries, without medical records, were not adjudicated, potentially leading to an underestimation of all-cause mortality rates.
  • Unadjudicated subsequent strokes: Subsequent hospitalized strokes were not adjudicated, which could affect the accuracy of subsequent stroke and mortality rates, as well as subtype classification.

Future work could focus on refining ML models with more comprehensive data, exploring the impact of regional variations on diagnostic practices and treatment, and investigating the underestimation of CE subtypes due to limited cardiac monitoring. The study also suggests that these ML-based classifications could improve IS etiology and prognosis assessment in population studies and pave the way for personalized treatment strategies.

Summary updated

Abstract

This study highlighted substantial heterogeneity in prognosis of IS subtypes and utility of ML approaches for classification of IS cases with incomplete clinical investigations.
BibTeX
@article{chun2023heterogeneity,
  author = {Matthew Chun and Haiqiang Qin and Iain Turnbull and Sam Sansome and Simon Gilbert and Alex Hacker and Neil Wright and Tingting Zhu and David A. Clifton and Derrick Bennett and Yu Guo and Pei Pei and Jun Lv and Canqing Yu and Ling Yang and Liming Li and Yan Lu and Zhengming Chen and Benjamin J. Cairns and Yiping Chen and Robert Clarke},
  title = {{Heterogeneity in the diagnosis and prognosis of ischemic stroke subtypes: 9-year follow-up of 22,000 cases in Chinese adults}},
  journal = {International Journal of Stroke},
  year = {2023},
  doi = {10.1177/17474930231162265},
}
Figure