Deep learning models may become a useful tool for determining the risk of relapses in patients with neuromyelitis optica spectrum disorder (NMOSD), according to a study recently published in Frontiers in Neurology.
“This study confirmed the superiority of deep learning to design a prediction model of relapse in patients with [aquaporin-4 antibody (AQP4-ab)]-positive NMOSD, with the LogisticHazard model showing the best predictive power in validation,” the authors wrote.
This retrospective study included 358 patients previously diagnosed with NMOSD from 1 center who underwent 1135 treatment episodes as the training cohort. The validation set included 92 patients who had received 213 treatment episodes from 9 different centers. All participants had AQP4-ab-positive NMOSD.
The researchers selected 9 variables: sex, disease duration, age at treatment initiation, maintenance therapy, previous acute attack under the same treatment, phenotype of most recent acute attack, annualized relapse rate of the previous year using the concordance index (C-index), AQP4-ab titer, and the Expanded Disability Status Scale (EDSS) score upon treatment initiation.
Read more about NMOSD complications
Random survival forest (RSF), LogisticHazard, DeepHit, and DeepSurv were the 4 deep learning models the study compared against the conventional Cox proportional hazard (COX-PH) to predict relapses using the aforementioned items.
The validation phase determined that the C-index produced by the LogisticHazard model (0.718) was superior to all others. Next came DeepHit with 0.704, DeepSurv with 0.698, RSF with 0.685, and lastly, Cox-PH at 0.651.
In the training set, RSF outperformed all others, with a C-index of 0.739. DeepHit was second, followed by LogisticHazard, DeepSurv, and Cox-PH with C-indices of 0.737, 0.722, 0.698, and 0.679, respectively.
These overall findings represent promising advances for patients with NMOSD since they translate into a potential risk stratification system that could allow healthcare providers to select which patients would benefit from certain therapies and which would not, hence diminishing unnecessary interventions.
“Based on the above-optimized model, a personalized treatment recommender system could be developed to minimize the probability of relapse in the future,” the authors concluded.
Wang L, Du L, Li Q, et al; Pan-Yangtze River Delta Alliance for Demyelinating Disease. Deep learning-based relapse prediction of neuromyelitis optica spectrum disorder with anti-aquaporin-4 antibody. Front Neurol. Published online August 5, 2022. doi:10.3389/fneur.2022.947974