Machine learning models comprising clinical and demographic features accurately predicted neurofibromatosis type 1 (NF1)-associated clinical features, according to study findings published in Neurology Clinical Practice.

Study researchers extracted electronic health record data and clinical registry information from data within the Washington University NF Clinical Program (collected 2002-2016). They incorporated a total of 27 unique clinical features and longitudinal data about NF1 development in the models. The researchers utilized a machine learning approach to predict the development of NF1-associated clinical features.

The NF1 clinical registry compared patients (n=798) who were 47.7% girls or women, and who had a mean age 13.0 (standard deviation [SD], 11.0) years. 81.1% of patients were White.


Continue Reading

Among the pediatric population (n=578), more boys presented with attention-deficit hyperactivity disorder (40.3% vs 26.3%; P <.001), girls with scoliosis (25.2% vs 16.0%; P =.01), non-White patients with skinfold freckling (97.2% vs 90.9%; P =.04), and White patients with Lisch nodules (63.6% vs 50.0%; P =.009), optic pathway glioma (20.9% vs 10.6%; P =.02), and T2-hyperintenisty basal ganglia (25.3% vs 14.2%; P =.02) and cerebellum (25.7% vs 13.3%; P =.01).

Children in the NF1 cohort were more likely to present with a maternal family history of NF1 (28.3%) compared with a paternal history (17.3%; X2, 15.5; P <.001).

With the clinical and demographic features and electronic health record data, the study researchers best predicted the onset of optic pathway glioma with an area under the receiving operator characteristic curve (AUC) of 0.82 (SD, ±0.06), sensitivity of 0.78 (SD, ±0.06), specificity of 0.78 (SD, ±0.07), and positive predictive value (PPV) of 0.78 (SD, ±0.05).

The model was second best at predicting the onset of attention-deficit hyperactivity disorder with an AUC of 0.74 (SD, ±0.05), sensitivity of 0.67 (SD, ±0.05), specificity of 0.68 (SD, ±0.08), and PPV of 0.68 (SD, ±0.06). The model performed poorest for the prediction of plexiform neurofibromas (AUC, 0.69±0.08; sensitivity, 0.62±0.07; specificity, 0.66±0.09; PPV, 0.65±0.1).

This study was based on electronic health records and may have included bias and inaccurate or missing data.

These data indicated that clinical and demographic characteristics could be used in models to predict the onset of 3 clinical features typically diagnosed among patients with NF1. Study researchers concluded, “Naïve machine learning techniques can be potentially used to develop and validate predictive phenotype complexes applicable to risk stratification and disease management in NF1.”

Reference

Morris SM, Gupta A, Kim S, Foraker RE, Gutmann DH, Payne PRO. Predictive Modeling for Clinical Features Associated with Neurofibromatosis Type 1. Neurol Clin Pract. Published online April 14, 2021. doi:10.1212/CPJ.0000000000001089