Machine learning based quantitative pain assessment for the perioperative period

Fig. 1. ROC curve of the pain assessment model. a ROC curve and model performances of intraoperative pain assessment, b ROC curve and model performances of postoperative pain assessment. ‘+’ represents the optimal cut-off. ROC receiver operating characteristic, AUROC area under the ROC curve, SPI surgical pleth index, SE sensitivity, SP specificity.

Fig. 2. Distributions of pain assessment model output. a Distribution of perioperative model’s output in intraoperative conditions, b distribution of SPI in intraoperative conditions, c distribution of perioperative model’s output in postoperative conditions, d distribution of SPI in postoperative conditions. In the intraoperative condition, the median deviation of the output of perioperative model for no pain and pain was the same with the deviation of SPI (deviation = 31). On the other hands, in the postoperative condition, the output of perioperative model’s median deviation between pain and no pain is 39, which is larger than SPI’s deviation, 7. SPI surgical pleth index.

Fig. 3. Feature importance represented by the SHAP value. PAsys has a larger value in the absence of pain and decreases with the occurrence of pain, and Skew has a similar trend to PAsys. AVonset and PW50/PPIonset increased with pain. In the absence of pain, the variation in SHAP values for AVonset was small. PAsys area of systolic phase in a pulse waveform, Skew skewness of a pulse waveform, AVonset amplitude variation of adjacent pulse onsets, PW50 pulse width at 50% of maximum amplitude, PPIonset pulse-to-pulse interval calculated with pulse onsets, SHAP Shapley additive explanations.

Fig. 4. Pain-induced photoplethysmogram waveform changes. a Intra-pulse waveform change according to the pain stimuli, b inter-pulse waveform change according to the pain stimuli. In response to pain, PPG waveform changes are observed, with a notable decrease in PPG amplitude. Additionally, the pulse width decreases in the presence of pain, which is primarily associated with a reduction in the pulse width during the diastolic phase (a). The overall baseline of the PPG exhibits significant variability when pain is present (b). PPG photoplethysmogram.

Fig. 5. Overall data collection procedure and analysis intervals. Data were collected throughout the patient’s entire surgical journey, including the preoperative, intraoperative, and postoperative periods. The analysis intervals were determined to be 2 min before and after pain stimulation.

Fig. 6. Characteristics and features of the photoplethysmographic waveform. a raw waveform-based characteristics. b triangulated waveform-based features. c inter-pulse-based features. A amplitude, ACA AC amplitude, AV amplitude variation, L length, PA pulse area, PPI pulse-to-pulse interval, PW pulse width, RS rising slope, FS falling slope, TPA triangulated pulse area, sys systolic, dia diastolic.

Fig. 7. Detailed process for feature selection and model development. The development process of the pain assessment model is divided into two main stages: “feature selection” and “model training and validation.” The first stage involves selecting that best features that reflect pain for both intraoperative and postoperative pain assessment models from a pool of all features. Feature importance is determined using SHAP values, and the machine learning algorithm is used for classification. In the model training and validation stage, the dataset is split into a training set and a validation set in an 8:2 ratio. To account for variability in feature importance, the training process is repeated 10,000 times with random splits, and the average SHAP ranking is calculated. To develop a perioperative pain assessment model, the top two features from each of the intraoperative and postoperative models are selected and integrated to develop a perioperative model. Modal development and validation are performed using 5-fold cross-validation. SHAP Shapley additive explanations.