Table 3 Machine Learning algorithms used for pain assessment grouped by type of pain assessed and algorithm classification
From: Pain assessment using physiological responses/markers in different types of pain: a scoping review
Type of Pain | Algorithm implemented | Pain Assessment Approach | Physiological Markers | Outcomes | Author, Year | |
|---|---|---|---|---|---|---|
Acute | SL / DL | BiLSTM-XGB | (M) Pain intensity (4 levels) | EDA. SCL (mean, SD, range, AUC) and SCR (mean, SD, range, peaks max, peaks min, peaks sum, peaks num, duration mean, slope mean, AUC) from cvxEDA, whole signals extracted | BL vs P3. Precision 0.87, recall 0.84, F1 0.86 | Pouromran et al., 202265 |
Shallow Learning (SL) | SVR | (M) Pain intensity (scale) | EDA. Time interval between successive extreme events above the mean and below the mean, and exponential fit to successive distances in 2-dimensional embedding space | MAE 0.93 ± 0.21. RMSE 1.16 ± 0.23 | Pouromran et al., 202167 | |
(M) Stimulation intensity (5 levels) | fNIRS. 10 statistical features from ΔHbO and ΔHHB each | Accuracy 66.55%, sensitivity 93.8%, specificity 96.14, F1 score 96.98% | Khan et al., 202480 | |||
SVM | (C) Pain detection | EDA. range, std, max, q3, sum, mean, median, q1, lqr, min | Accuracy 93.2% ± 8, precision 94.6% ± 6, sensitivity 96.8% ± 3 F1-score 95.5% ± 3. | Fernandez-Rojas et al., 202357 | ||
(C) Pain intensity (4 levels) | EMG. Corrugator amplitude peak-to-peak, corrugator Shannon entropy. ECG.HRV R-R slope | BL vs P3. Accuracy 77.05% | Walter et al., 201424 | |||
(C) Pain intensity (4 levels) | EMG. Zygomaticus (similarity correlation, SD of mean vector, amplitude RMS, linearity lag dependence, variability variance), corrugator and trapezius (amplitude peak, similarity correlation, similarity mutual information) | BL vs P3. Accuracy 90.94%, sensitivity 92.24%, specificity 89.65% | Gruss et al., 201525 | |||
(M) Pain intensity (4 levels). | Not reported | Accuracy 81.39% | Chu et al., 201785 | |||
RFc | (C) Pain intensity (5 levels) | EDA. SCR, SCL, dPhEDA, TVSymp (Max, min, range, SD, interquartile range, mean, local maxima and minima for all) | BL vs P4: Volunteers. Accuracy 91.70% Patients. Accuracy 89.58% | Khan et al., 202382 | ||
(C) Stimuli detection | EMG. Reflex withdrawal (Ipsilateral and Contralateral Limb). ECG. HR. PPG. SPO2. EEG. Noxious-evoked brain activity. Facial expression | Accuracy 0.81 (0.70–0.89), AUC: 0.9 (0.78–0.95). | Van der Vaart., 201986 | |||
(M) Pain intensity (5 levels). | EDA. Mean, max, RMS, IQR, mean maxima | BL vs P1. Accuracy 86% | Aqajari et al., 202164 | |||
(C) Stimulus intensity (5 levels) | ECG, EDA, EMG. 155 feat. including amplitude, variability, stationarity, entropy, linearity, similarity and freq. | BL vs P4. Accuracy 92% | Albahdal et al., 202487 | |||
Logistic regression | (C) Stimulus intensity (5 levels) | EMG (zygomaticus). FF, ZCD. EDA. SCR (slope) | BL vs P4. Accuracy 80%. Sensitivity 78%. Specificity 83% | Sen & Pal, 2021125 | ||
(M) Pain intensity (4 levels) | ECG. HRV (RRI, LF, HF). PPG. PPGA, ANSS, BL | Accuracy 50%, sensitivity 60%, specificity 72% | Jhang et al., 202131 | |||
LightGBM | (C) Pain detection | ECG. Frequency spectrum analysis | Accuracy 0.9. Specificity 0.91. Sensitivity 0.9 Precision 0.87. F1-Score 0.88. AUC 0.96. | Chu et al., 202491 | ||
Bi-layered NNs | (C) Pain intensity (3 levels) | PPG. Time domain and PRV features | BL vs P2. 69% accuracy, 83.33% sensitivity and 75% specificity. | Khan et al., 202393 | ||
DL | Transformer-based Architecture | (C) Pain detection | ECG+FaceExp | Accuracy 82.74% | Gkikas et al., 2024121 | |
TCAtt-PainNet | (C) Pain detection & intensity (5 levels) | EDA. 11 statistical features. ECG. HR, HRV (SDNN, RMSSD) | MAE 0.87 ± 0.19. RMSE 1.07 ± 0.21. R2 0.39 ± 0.24. ICC 0.56 ± 0.22 | Jiang et al., 2024101 | ||
Deep Learning (DL) | Personalised Module (SensMeasure) + Base NN with DynAtt | (C) Pain intensity (5 levels) | EDA. Max, average, AUC and SD of SCR and SCL from cvxEDA. Phasic driver (average rise time, number of peaks, max peak amplitude). ECG. HR(mean, median), HRV (SDNN, RMSSD, LF). Sensitivity. SensMeasure | P0 vs P4. MAE 0.93 ± 0.18. RMSE 1.12 ± 0.20 R^2 0.34 ± 0.23 ICC 0.50 ± 0.22. | Jiang et al., 202436 | |
Parallel TCN-SBU-LSTM | (C) Pain detection | EDA. SCR, TFS-phEDA | Accuracy 93.1%, F1-score 77.8%, AUC 0.967, sensitivity 82.3%, specificity 95% | Pinzon-Arenas et al., 202366 | ||
MLPNN | (M) Stimulus intensity (3 levels) | EDA. SCR, SCL, and dPhEDA obtained using sparsEDA. TVSymp and MTVSymp | Accuracy 69.7%, R2 0.357, M-RMSE 0.936, M-MAE 0.762 | Posada-Quintero et al., 202161 | ||
LSTM-SW with DF | (M) Pain intensity (4 levels) | EDA. 1st & 2nd derivative, min, max, mean, SD. | ICC 0.33, MSE 0.08 | Othman et al., 202396 | ||
Late fusion CNNs | (C) Pain intensity (4 levels) | EDA, EMG (trapezius), ECG | BL vs P4. Accuracy 84.40 ± 14.43 % | Thiam et al., 201998 | ||
Kernel ELM-based classifier | (M) Pain intensity (4 levels) | EEG. Alpha band | Accuracy 68.90 ± 3.12 % | Yu et al,. 202094 | ||
DFB-ConvNets | (M) Pain intensity (3 levels) | EEG. Alpha, Gamma, and Beta bands. | Accuracy 97.37%, precision 96.05%, specificity 98.03%, sensitivity 96.06%, F1-Score 96.05% | Yu et al., 202099 | ||
Deep-Attentive-Recurrent-CNN | (M) Pain intensity (4 levels) | EEG. Spatial spectral-temporal signal analysis | Accuracy 92.14 %, F1 score 92.11 % | May et al., 2021117 | ||
CNN_LSTM | (C) Pain intensity (4 levels) | ECG, EDA. 2D waveform | NP-PL4. Accuracy 94.12%. MSE 0.0588 | Subramaniam & Dass, 2021100 | ||
Acute/Chronic | DL | LSMT | (M) Pain intensity (3 levels) | EDA. SCR and SCL (median, SD, min, max 1st and 2nd derivative) | Accuracy 0.89 ± 0.05, F1-Score 0.85 ± 0.06 | Badura et al., 2024109 |
Chronic | SL | RFc | (M) Pain intensity (4 levels) | PPG. PPGA number of changes, PRV (PR min, PR max, PR median, SDNNI, SDANN, RMSSD) | Accuracy 0.768 ± 0.012. Specificity 0.869% ± 0.007, sensitivity 0.737 ± 0.016, F1-score 0.768 ± 0.012. | Patterson et al., 202388 |
Modified K-Means | (C) Brain microstate analysis (5 levels) | EEG. GFP peaks | N/A | May et al., 2021 | ||
Perioperative | DL | MLPNN | (M) Pain intensity | ECG. ANI | MAE 2.848 ± 0.308 | Jean et al., 202297 |
Postoperative | SL | Logistic Regression | (C) Pain detection | PPG. NPI | Accuracy 75.3%, AUC: 0.8018 | Cho et al., 201853 |
(C) Pain detection | PPG. RMSSD-ACVonset/ACAbl, AV-Asys/Atotal, SD-RS | Accuracy 0.752, AUC 0.825 | Seok et al., 201984 | |||
(C) Pain detection | PPG. ACVsys/ACAdia | Accuracy 79.50%, sensitivity 74.02%, specificity 85.99% | Yang et al., 201847 | |||
Linear SVM + KDE + weighted Bayesian fusion | (C) Pain detection | EDA. TSD triangular matrix (mean, SD, entropy). SCL (SD). Video. 20 discrete Aus, three head pose indicators. | Accuracy 90.91%, sensitivity 100%, specificity 81.82% | Susam et al., 202292 | ||
DL | DBN (Selective bagging) | (C) Pain detection (4 levels) | PPG. Pulse height, rise time, fall time, average HR, PRV (HF, LF, VLF, LF/HF, AVNN, SDNN, RMSSD, NN20, pNN20, NN50, pNN50) | Accuracy 86.79 %, AUC 0.841 ± 0.039 | Lim, 201956 | |
CNN | (C) Pain detection | PPG. 2D spectrogram. SPI | Accuracy 71%, sensitivity 68.1%, specificity 73.8%, AUC 0.757 | Choi et al., 2021107 | ||
