Deep learning can help predict functional thrombolysis outcomes

Deep learning (DL) trained to read CT brain scans can help clinicians predict functional thrombolysis outcomes, according to an April 30 study published in Academic Radiology.

“…thrombolysis exposes patients to significant risks, including intracranial hemorrhage and gastrointestinal bleeding,” wrote Stephen Bacchi, MBBS, of Royal Adelaide Hospital in Australia, and colleagues. “Therefore, the appropriate selection of patients for thrombolysis is vitally important, and additional methods to aid in the selection of such patients is an ongoing clinical need.”

The researchers collected noncontrast CT brains scans from 204 consecutive patients who presented with an acute stroke and received subsequent intravenous thrombolysis across two hospitals from 2009 to 2015.

Convolutional neural network (CNN) and artificial neural network (ANN) models were created to predict either improvements in the National Institutes of Health Stroke Scale of ≥4 points at 24 hours (NIHSS24) or modified Rankin Scale 0-1 at 90 days (mRS90).

For predicting NIHSS24, the top model was a combined CNN and ANN approach which achieved an accuracy of 0.71, AUC of 0.70 and F1 score of 0.69. For NIHSS24 prediction, the CNN and ANN combination was also the most effective, achieving an accuracy of .074, AUC of 0.75 and an F1 score of 0.69.

“The results of this study demonstrate proof of the concept that DL models may aid in the prediction of thrombolysis outcomes,” the authors wrote.

Bacchi and colleagues noted patients with unfavorable noncontrast CT imaging were likely not treated with thrombolysis, “limiting discriminative ability.” Future models would likely improve with CT angiogram and/or CT perfusion images.

The researchers suggested forecasting clinical outcomes could become the “gold standard” in applying DL to medical conditions.

“This is the case because, if assessed objectively, future clinical outcome may be a clinically useful, reliable ground-truth that may be obtained noninvasively.”

Bacchi et al. also believe deep learning may be able to predict more than just functional thrombolysis outcomes.

“It would be additionally worthwhile testing this approach in outcome prediction following thrombectomy,” they wrote. “Larger datasets may allow the prediction of other clinically significant thrombolysis outcomes with DL models, in particular symptomatic intracranial hemorrhage.”