Deep Learning Algorithm Shows High Accuracy in Predicting ICH Hematoma Expansion
The artificial neural network demonstrated an area under the precision-recall curve of 0.92 with recall of 0.72 with precision of 0.39 during testing, all values that were greater than classic regression models.
Arooshi Kumar, MD
Relative to traditional logistic regression models (GLM), a newly developed deep learning algorithm, an artificial neural network (ANN), demonstrated improved sensitivity and similar positive predictive value in predicting hematoma expansion (HE) in patients with intracerebral hemorrhage (ICH).1
The data, presented at the 2022 International Stroke Conference (ISC), February 9-11, in New Orleans, Louisiana, stems from the ATACH 2 trial (NCT01176565), which originally was a study that evaluated systolic blood pressure (SBP) when treating acute hypertensive response in patients with ICH. In this exploratory analysis, a multilayer feedforward ANN was trained using the back-propagation method to minimize the loss function with 5-fold validation.
Lead author Arooshi Kumar, MD, neurology fellow, Penn Medicine, University of Pennsylvania, and colleagues, defined HE as an increase in hematoma volume of either 33% or more than 6 mL within the first 24 hours. They compared the predictability value of the ANN to GLM in a cohort of 963 patients (mean age, 62 years [±13.0]; 38.5% female) with ICH by calculating area under the precision-recall curve (AUPRC), recall, and precision. Of the cohort, 80% of the patients were used for training and 20% were for testing.
READ MORE: NeuroVoices: Imad Najm, MD, on Aiming to Prevent Neurological Disorders Before Symptom Onset
Of the cohort, 31% demonstrated HE, with an initial recorded median hematoma volume of 10.5 mm3 (IQR, 5.16-20.25). Additionally, these patients had median admission SBP of 200 (IQR, 184-217), median platelet count of 213 (IQR, 178-256), median international normalized ratio (INR) of 1.0 (IQR, 0.9-1.0) and median admission Glasgow Coma Scale score of 15 (IQR, 13-15). After applying GLM models in the testing cohort, investigators recorded AUPRC of 0.38 with a recall of 0.60 and precision of 0.38. In comparison, the ANN training model demonstrated AUPRC of 0.92 with recall of 0.72 with precision of 0.39 during testing.
"We support that ANN can capture complex associations between variables not attainable in classic regression models,” Kumar et al concluded. "This model may help identify patients at risk for HE who warrant careful monitoring and aggressive treatment upfront including those suitable for clinical trials for treatments."
ATACH-2 was originally a large-scale study that randomized eligible participants with ICH to an SBP target of 110-139 mm Hg (intensive treatment) or a target of 140-179 mm Hg (standard treatment) in order to test the superiority of intensive reduction of SBP to standard reduction. Investigators administered intravenous nicardipine within 4.5 hours after symptom onset to lower blood pressure.2
The primary outcome, death or disability, was observed in 38.7% (186 of 481) of the participants in the intensive-treatment group and 37.7% (181 of 480) in the standard-treatment group (relative risk, 1.04; 95% CI, 0.85-1.2). These findings suggested that intensive reduction in the SBP level does not provide an incremental clinical benefit.
For more coverage of ISC 2022, click here.
