Predicting 90-Day Outcomes in Acute Ischemic Stroke from Baseline Non-Contrast CT: A Deep Learning Approach vs. Alberta Stroke Programme Early CT Score
Abstract Body: Purpose: Predicting long-term clinical outcomes based on immediate hospital admission data is invaluable for treatment decisions, prognostication, resource management, clinical trials, and setting patient expectations. Non-Contrast CT (NCCT) scans are fast, widely available, and can detect early stroke signs. However, the current interpretation method, the Alberta Stroke Program Early CT Score (ASPECTS), is subjective and may have limited ability to detect subtle changes or complex presentations. This study aimed to predict functional outcomes in acute ischemic stroke (AIS) patients using deep learning (DL) from the initial NCCT and compare its performance with the ASPECTS.
Methods: A total of 1,365 AIS patients from Lausanne University Hospital, who had baseline non-contrast CT scans and 90-day mRS data, were randomly divided into three groups: 70% (n=951) for model training, 15% (n=205) for validation, and 15% (n=205) for independent testing. The DL model utilized a tailored 3D Residual neural network structure, with the neuron count in the final fully connected layer reduced to one, enabling the model to generate a continuous prediction of the 90-day modified Rankin Scale (mRS) score. For ASPECTS prediction, a Support Vector Regressor was used to also generate continuous predictions of the 90-day mRS score. Evaluation metrics focused on the determination of unfavorable outcomes, defined as mRS scores greater than 2, comparing the performance of both the ASPECTS and the DL model.
Results: In the testing set, to predict unfavorable outcomes, the DL model demonstrated enhanced performance with an AUC of 0.74 (95% CI: 0.67, 0.81) compared to ASPECTS' AUC of 0.61 (95% CI: 0.53, 0.69; p = .01). This improvement was particularly notable in patients with unknown reperfusion status and those without EVT, where the DL model achieved significantly higher AUCs of 0.77 (95% CI: 0.66, 0.87; p=.03) and 0.75 (95% CI: 0.68, 0.83; p=.008), respectively, compared to ASPECTS' AUCs of 0.59 in both categories.
Conclusions: The proposed DL model outperformed the conventional ASPECTS in predicting unfavorable outcomes from baseline non-contrast CT scans.
Liu, Yongkai
( Stanford University
, Stanford
, California
, United States
)
Van Voorst, Henk
( Stanford University
, Stanford
, California
, United States
)
Jiang, Bin
( Stanford University
, Stanford
, California
, United States
)
Yu, Yannan
( Stanford University
, Stanford
, California
, United States
)
Zaharchuk, Greg
( Stanford University
, Stanford
, California
, United States
)
Author Disclosures:
Yongkai Liu:DO NOT have relevant financial relationships
| Henk van Voorst:DO have relevant financial relationships
;
Research Funding (PI or named investigator):Dutch Scientific Council (NWO):Active (exists now)
; Research Funding (PI or named investigator):Dutch Heart Foundation:Past (completed)
; Royalties/Patent Beneficiary:Stanford School of Medicine:Active (exists now)
; Researcher:Amsterdam UMC:Past (completed)
| Bin Jiang:DO NOT have relevant financial relationships
| Yannan Yu:DO NOT have relevant financial relationships
| Greg Zaharchuk:DO have relevant financial relationships
;
Ownership Interest:Subtle Medical Inc.:Active (exists now)
