As our results indicated that NR and LR neurons have very different electrophysiological properties, we next wondered if we could predict the patient’s response to Li based only on electrophysiological measurements from its derived neurons. To make these findings applicable for medical purposes, we separated the data into training data composed of recordings of five patients (see Figure 4a), and then classified recordings from a patient who was new to the model (see Figure 4b). We used an NB classifier (see Materials and Methods), which learns the posterior probabilities of features extracted from the electrophysiological measurements given that the patient was an NR or LR patient. This model was then used to classify the remaining patient (see Figure 4b). Performance was assessed by repeating this process 2000 times and the patient left out of the training data was a different patient each time. Since a patch-clamp recording session usually consists of patching a few neurons, to improve performance, we used, as the test data, 1-, 3- or 5-cell recordings (from 1 or 3 or 5 cells accordingly, see Materials and
