Outcome prediction comparison of ischaemic areas' radiomics in acute anterior circulation non-lacunar infarction

Xiang Zhou; Jinxi Meng; Kangwei Zhang; Hui Zheng; Qian Xi; Yifeng Peng; Xiaowen Xu; Jianjun Gu; Qing Xia; Lai Wei; Peijun Wang

doi:10.1093/braincomms/fcae393

Outcome prediction comparison of ischaemic areas' radiomics in acute anterior circulation non-lacunar infarction

Brain Commun. 2024 Nov 15;6(6):fcae393. doi: 10.1093/braincomms/fcae393. eCollection 2024.

Authors

Xiang Zhou¹, Jinxi Meng¹, Kangwei Zhang¹, Hui Zheng², Qian Xi³, Yifeng Peng⁴, Xiaowen Xu¹, Jianjun Gu⁵, Qing Xia⁶, Lai Wei¹, Peijun Wang^{1

7}

Affiliations

¹ Department of Radiology, Tongji Hospital, School of Medicine, Tongji University, Shanghai 200065, China.
² Department of Radiology, Xinhua Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai 200092, China.
³ Department of Radiology, Shanghai East Hospital, School of Medicine, Tongji University, Shanghai 200120, China.
⁴ Department of Radiology, Putuo Hospital, Shanghai University of Traditional Chinese Medicine, Shanghai 200062, China.
⁵ Department of Radiology, Shanghai General Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai 200080, China.
⁶ SenseTime Research, Shanghai 200232, China.
⁷ Institute of Medical Imaging Artificial Intelligence, Tongji University School of Medicine, Shanghai 200065, China.

Abstract

The outcome prediction of acute anterior circulation non-lacunar infarction (AACNLI) is important for the precise clinical treatment of this disease. However, the accuracy of prognosis prediction is still limited. This study aims to develop and compare machine learning models based on MRI radiomics of multiple ischaemic-related areas for prognostic prediction in AACNLI. This retrospective multicentre study consecutively included 372 AACNLI patients receiving MRI examinations and conventional therapy between October 2020 and February 2023. These were grouped into training set, internal test set and external test set. MRI radiomics features were extracted from the mask diffusion-weighted imaging, mask apparent diffusion coefficient (ADC) and mask ADC620 by AACNLI segmentations. Grid search parameter tuning was performed on 12 feature selection and 9 machine learning algorithms, and algorithm combinations with the smallest rank-sum of area under the curve (AUC) was selected for model construction. The performances of all models were evaluated in the internal and external test sets. The AUC of radiomics model was larger than that of non-radiomics model with the same machine learning algorithm in the three mask types. The radiomics model using least absolute shrinkage and selection operator-random forest algorithm combination gained the smallest AUC rank-sum among all the algorithm combinations. The AUC of the model with ADC620 was 0.98 in the internal test set and 0.91 in the external test set, and the weighted average AUC in the three sets was 0.96, the largest among three mask types. The Shapley additive explanations values of the maximum of National Institute of Health Stroke Scale score within 7 days from onset (7-d NIHSS_max), stroke-associated pneumonia and admission Glasgow coma scale score ranked top three among the features in AACNLI outcome prediction. In conclusion, the random forest model with mask ADC620 can accurately predict the AACNLI outcome and reveal the risk factors leading to the poor prognosis.

Keywords: acute ischaemic stroke; magnetic resonance imaging; model construction; prognosis prediction; radiomics.