Machine Learning-Based Model for Individualized Drug Dose Prediction for Propofol

Study Methods:

（1）Case selection： Patients requiring elective surgical treatment in the Cardiovascular and Cerebrovascular Disease Hospital of General Hospital of Ningxia Medical University.

Inclusion criteria: Patients were gender-neutral and aged ≥18 years.

（2) Clinical study protocol： Patients were admitted to the operating room to establish intravenous access, connected to ECG monitoring, EEG monitoring, 4L/min mask oxygenation, and real-time video recording of the entire anesthesia induction process using a video recorder for postoperative integration of various data. Propofol was pumped in at 100 mg/kg/h, and the anesthesiologist with propofol assessed the degree of sedation of the patient until the patient was deeply sedated and the MOAA/S score was 0, and the pumping of propofol was stopped.

(3) Observation indicators： Demographic information and general preoperative data were collected, including patients' gender, age, weight, height, ASA classification, body mass index (BMI), and past medical history (hypertension, coronary heart disease, diabetes mellitus, respiratory disease, stroke).

Systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, finger pulse oximetry (SPO2), electroencephalogram waveforms, the dosage of propofol administered from the start of anesthesia induction to MOAA/S = 0, and the time of administration of propofol were recorded for all the basic and extended monitoring parameters, respectively, after the patient was admitted to the operating room and at the time of the 0 score of the MOAA/S score.

(4) Evaluation methods of each item Depth of anesthesia assessment: MOAA/S depth of sedation was used for assessment, and MOAA/S score of 4-5 was classified as mild sedation, which indicated that the response to calling names in normal tone was sensitive or slow; 2-3 was classified as moderate sedation, which indicated that there was a response to calling names loudly or repeatedly, or there was a response to slight pushing and vibration; ≤1 was classified as deep sedation, which indicated that there was a response to pain stimulation; 0 was classified as general anesthesia, which indicated that there was a response to pain stimulation; 0 was classified as general anesthesia, which indicated that there was a response to pain stimulation. ≤1 is classified as deep sedation, indicating a response to painful stimuli; 0 is classified as general anesthesia, indicating no response to painful stimuli.

EEG feature extraction: use python MNE (https://mne.tools/stable/index.html) module package for analysis, MNE can read most common physiological signals in raw data format, the specific methods are as follows: ① Data extraction: import data → electrode positioning → reject useless electrodes → re-reference → filter → drop sampling rate → run ICA → batch processing → segmentation and baseline correction → interpolation of bad conductance and rejection of bad conductance → removal of noise components → rejection of bad segments; ② Data division: the data of the required EEG monitoring points are divided to include wakefulness and sedation; ③ Feature extraction: time domain, amplitude square root value, kurtosis, skewness, burst suppression rate/minute, frequency domain, and entropy index are obtained by Fourier transform.

(5) Proposed machine learning algorithm: regression analysis is a predictive modeling technique that works from a set of data to determine quantitative relational equations between certain variables, statistically test these relational equations between the variables, and identify the variables that have a significant effect from among the multiple variables that affect a particular variable.