Epilepsy is largely under-diagnosed in low-income and middle-income countries, due to lack of medical specialists and expensive electroencephalography (EEG) hardware. In this study we investigate if low-cost consumer-grade EEG in combination with machine learning techniques can offer a reliable screening tool to improve diagnosis rates. We acquired brain signals in people with epilepsy (N=163) and healthy controls (N=138) in two difficult-to-reach areas in rural Guinea-Bissau and Nigeria. Five minutes of fourteen channel resting-state EEG data were acquired with a portable, low-cost consumer-grade EEG recording headset. EEG channel time-series were divided in four-second artifact-free epochs and transformed into delta, theta, alpha, beta and gamma wavelet frequencies. Summary measures such as the mean, standard deviation, minimal value, maximal value of the epoch signal fluctuations were used to train a random forest classifier. Epilepsy diagnosis based on at least three months seizure calendar data was used as the gold standard diagnosis. To prevent too optimistic classification the trained model was evaluated with EEG data from subjects not used in the training. In addition, we tested a classification model trained on Nigeria data against data from people in Guinea-Bissau and vice versa. The most contributing data features in the EEG were found in the beta and theta frequencies in Guinea-Bissau and Nigeria, respectively. Within-country model performance was good with area under the receiver-operating curves of 0.85 and 0.78 (± 0.02 standard errors) in unseen data in Guinea-Bissau and Nigeria, respectively. Across-country performance was moderate (0.62 and 0.64 ± 0.02). Our data suggests that a combination of low cost electroencephalography and machine learning techniques may facilitate diagnostic screening for epilepsy in the most remote areas of the world.

Reliable and automatic epilepsy classification with affordable, consumer-grade electroencephalography in rural sub-Saharan Africa