Extended Data Fig. 4: Classifier models and their performance in classifying neurons as recorded from high- or low-grade glioma.

a, Two-dimensional projection of training data using DFAC illustrating the separation of high- (blue) and low-grade (green) recordings, in the optimised discriminant (dimension-reduced) space. Performance for one classification test, out of 100 tests with random training data is shown. Large data points represent the centroids for high- (blue) and low-grade (green). b, DFACf projected discriminant scores for both training and test data for the sample shown in (a) illustrating the separation of training and test sets. c, The architecture of the LRC presents that multiple input features (or trace values) are first combined linearly by assigned weights and an intercept (bias), forming a weighted sum, and then this output is transformed by the sigmoid (logistic) activation function, mapping the value to a probability between 0 and 1. A binary prediction can ultimately be made to the model’s confidence on distinguishing the classes. d, The architecture of MLPC depicts an input layer that receives the raw data, which then feeds into a single hidden layer comprising 100 neurons. Each neuron in this hidden layer applies a non-linear activation function (ReLU) to its weighted sum of inputs, whose outputs are then passed to an output layer, producing the final class prediction. e, The classification performance in training strategies (n = 100 tests), including those fed with extracted features (DFACf and LRCf) and raw data (LRCt and MLPCt) (One-way ANOVA). Data are presented as mean values ± SEM. f, The performance of DFAC against the number of neurons used for training. Linear regression, R2 = 0.8847. **, p < 0.01; ****, p < 0.0001.