Cancer Prediction With Machine Learning of Thrombi From Thrombectomy in Stroke: Multicenter Development and Validation

Joon Nyung Heo, Hyungwoo Lee, Young Seog, Sungeun Kim, Jang Hyun Baek, Hyungjong Park, Kwon Duk Seo, Gyu Sik Kim, Han Jin Cho, Minyoul Baik, Joonsang Yoo, Jinkwon Kim, Jun Lee, Yoonkyung Chang, Tae Jin Song, Jung Hwa Seo, Seong Hwan Ahn, Heow Won Lee, Il Kwon, Eunjeong ParkByung Moon Kim, Dong Joon Kim, Young Dae Kim, Hyo Suk Nam

Research output: Contribution to journalArticlepeer-review

3 Scopus citations


BACKGROUND: We aimed to develop and validate machine learning models to diagnose patients with ischemic stroke with cancer through the analysis of histopathologic images of thrombi obtained during endovascular thrombectomy. METHODS: This was a retrospective study using a prospective multicenter registry which enrolled consecutive patients with acute ischemic stroke from South Korea who underwent endovascular thrombectomy. This study included patients admitted between July 1, 2017 and December 31, 2021 from 6 academic university hospitals. Whole-slide scanning was performed for immunohistochemically stained thrombi. Machine learning models were developed using transfer learning with image slices as input to classify patients into 2 groups: cancer group or other determined cause group. The models were developed and internally validated using thrombi from patients of the primary center, and external validation was conducted in 5 centers. The model was also applied to patients with hidden cancer who were diagnosed with cancer within 1 month of their index stroke. RESULTS: The study included 70 561 images from 182 patients in both internal and external datasets (119 patients in internal and 63 in external). Machine learning models were developed for each immunohistochemical staining using antibodies against platelets, fibrin, and erythrocytes. The platelet model demonstrated consistently high accuracy in classifying patients with cancer, with area under the receiver operating characteristic curve of 0.986 (95% CI, 0.983-0.989) during training, 0.954 (95% CI, 0.937-0.972) during internal validation, and 0.949 (95% CI, 0.891-1.000) during external validation. When applied to patients with occult cancer, the model accurately predicted the presence of cancer with high probabilities ranging from 88.5% to 99.2%. CONCLUSIONS: Machine learning models may be used for prediction of cancer as the underlying cause or detection of occult cancer, using platelet-stained immunohistochemical slide images of thrombi obtained during endovascular thrombectomy.

Original languageEnglish
Pages (from-to)2105-2113
Number of pages9
Issue number8
StatePublished - 1 Aug 2023

Bibliographical note

Publisher Copyright:
© 2023 Lippincott Williams and Wilkins. All rights reserved.


  • machine learning
  • stroke
  • thrombectomy


Dive into the research topics of 'Cancer Prediction With Machine Learning of Thrombi From Thrombectomy in Stroke: Multicenter Development and Validation'. Together they form a unique fingerprint.

Cite this