The illustration above is an example of the NeuroPace® implantation. It includes the neurostimulator, attached to the skull, and depth and cortical leads, located inside the cranial cavity.

The illustration above is an example of the AspireSR® implantation for Vagus Nerve Stimulation (VNS). It includes the generator implanted in the upper-left chest and the lead of electrodes connected to the vagus nerve.

The intake data collected included identification features (name, database ID, medical record number) and demographic features (gender, year of birth). These were all found in the opening section of patient clinical reports and were easily identifiable.

The seizure experience data collected included timeline features (age of onset, age during surgery, year of surgery, and any prior surgeries) and seizure classification features (febrile seizures, status epilepticus, auras). The timeline features were found in the opening section of patient clinical reports, and the seizure classification features were found in the semiology summary of reports.

The neuroimaging and testing results collected included MRI, ictal EEG, interictal EEG, MEG, PET, and SPECT features. These were found in the Multidisciplinary Conference (MDC) summary of patient reports (see Index). The presurgical consensus reported the following: identified lesions and their locations under MRI results, regions of abnormal electrical activity detected during respective ictal and interictal periods under EEG results, number and location of dipoles detected under MEG results, locations of hypometabolism detected under PET scan results, and the difference between ictal and interictal regions of increased blood flow under SPECT scan results. Similar to how was done with SPECT scan results, the differences between the ictal and interictal EEG results were later determined using machine learning.

The RNS electrode data included count features (total electrode count, stimulating vs. nonstimulating counts), and electrode detail features (stimulating vs. nonstimulating types, stimulating locations, nonstimulating locations). These were found in the final section of patient reports detailing surgical plans. The stimulating electrodes were the one or two electrodes connected to the generator; the nonstimulating electrodes were any additional electrodes still implanted. The first part of the list of all electrodes used corresponded with the locations of stimulating electrodes; the second half of the list corresponded with the locations of unstimulating electrodes. For example, the only ds (depth-stimulating) electrode was implanted in the left temporal region, while the first d (depth-unstimulating) electrode was implanted in the left mesial temporal hippocampus region.

The success data collected included only the patient outcome dependent variable. The patient outcomes were determined to be either a lack of response to follow-up (-1), a failure (0), or a success (1). These determinations were subjectively made with the expertise of a medical professional but were based on objective factors such as reported improvements by a patient and any returns to antiseizure medication. Typically, such returns are only done if considerable improvement with the device have been exhibited.

The depiction above is the output from the machine learning program confirming the 25 features, each with 127 iterations, included in the dataset.

The coding input above produced a table that shows the counts of females (70) and males (57) whose reports were included in the dataset.

The histogram above shows the distribution of ages at which RNS surgery was undergone by the patients included in the dataset. A right skew and a peak in the early 20s are shown.

The depiction above is the output from the machine learning program confirming the counts of the respective outcomes, as well as a pie-chart plot of these outcomes to the immediate right.

Depicted above are two sets of 121-dimensional vector coordinates that respectively represent the word ‘frontal’ and the phrase ‘left mesial’ from an iteration of the ‘elec used’ feature. The latter coordinates were calculated by subtracting the coordinates associated with ‘frontal’ from those of ‘left mesial frontal’. This exemplified how Mathematica distinguishes between related terms that differ in small ways.

Depicted above are plotted values for the 90 iterations of the eegiToElecDistance and mriToElecDistance features, as well as the associated outcome response values. The mriToElecDistance values did not show a clear pattern in this plot, as the values were linked to the eegiToElecDistance values that were ordered from smallest to largest for each outcome value. A notably weaker slope of eegToElecDistance values was observed for successes compared to failures.

Depicted above is an ultra-simplified, 2D-vector space plot that represents ‘elec used’ feature values generated by Mathematica following dataset processing. Only the most general details regarding its structure can be visualized; however, a clear difference in aggregation can be seen for values beginning with ‘left’, ‘right’, and ‘bilateral’.

The depiction above is the input to the machine learning program reducing version 8 of the dataset to 10 features.

The depiction above is the input to the machine learning program reducing version 9 of the dataset to 9 features.

The LogisticRegression algorithm was used on the 8th version of the dataset to achieve a 70% (±15%) accuracy rate for predictions after using 80 examples for training and 10 examples for testing.

The NeuralNetwork algorithm was used on the 9th version of the dataset to achieve a 70% (±15%) accuracy rate for predictions after using 80 examples for training and 10 examples for testing.

The GradientBoostedTrees algorithm was used on the 9th version of the dataset to achieve a 70% (±15%) accuracy rate for predictions after using 80 examples for training and 10 examples for testing.

The odds ratio of machine accuracy to guessing accuracy was found using the low and high bounds of the program accuracy range, as well as the average value for guessing accuracy. This odds ratio was found to be 1.40 (±0.3), meaning one’s odds of being accurate with the CDSS in comparison to guessing were 40% (±30%) better.