Author Contributions
Conceptualisation, R.F.-R.; methodology, M.U.K., M.S. and R.F.-R.; formal analysis, M.U.K. and M.S.; investigation, M.U.K., M.S. and R.F.-R.; data interpretation: M.U.K., M.S. and R.F.-R.; resources, R.F.-R. and R.G.; data collection, N.H., C.J. and R.F.-R.; writing—original draft preparation, M.U.K. and M.S.; writing—review and editing, M.G., G.C., R.G. and R.F.-R.; supervision, R.F.-R. and R.G.; project administration, R.F.-R.; funding acquisition, R.F.-R. All authors have read and agreed to the published version of the manuscript.
Figure 1.
System block diagram of the proposed fNIRS-based pain assessment system.
Figure 1.
System block diagram of the proposed fNIRS-based pain assessment system.
Figure 2.
Schematic representation of the experimental procedure.
Figure 2.
Schematic representation of the experimental procedure.
Figure 3.
fNIRS channel information: (a) fNIRS cap. (b) Schematic of fNIRS channel locations. Red: Sources; Blue: Detectors; and Yellow: Channels. Specifically, the optodes Tx1, Tx2, Tx7, Tx9, Rx3, and Rx7 were positioned at the following locations on the standard 10–20 EEG system: Tx1: at F8; Tx2: at Fp2; Tx7: at Fp1; Tx9: at F7; Rx3: at F4; and Rx7: at F3.
Figure 3.
fNIRS channel information: (a) fNIRS cap. (b) Schematic of fNIRS channel locations. Red: Sources; Blue: Detectors; and Yellow: Channels. Specifically, the optodes Tx1, Tx2, Tx7, Tx9, Rx3, and Rx7 were positioned at the following locations on the standard 10–20 EEG system: Tx1: at F8; Tx2: at Fp2; Tx7: at Fp1; Tx9: at F7; Rx3: at F4; and Rx7: at F3.
Figure 4.
Twenty-two-Channel fNIRS (measuring changes in ) raw data (excluding two faulty channels) with annotated and highlighted durations for different conditions: B (Baseline), LA (Low Arm Pain), HA (High Arm Pain), HH (High Hand Pain), and LH (Low Hand Pain). The gray background in the figure represents the duration of each experiment phase: Baseline: 60 s, LA, LH, HA, and HH, each lasting 10 s.
Figure 4.
Twenty-two-Channel fNIRS (measuring changes in ) raw data (excluding two faulty channels) with annotated and highlighted durations for different conditions: B (Baseline), LA (Low Arm Pain), HA (High Arm Pain), HH (High Hand Pain), and LH (Low Hand Pain). The gray background in the figure represents the duration of each experiment phase: Baseline: 60 s, LA, LH, HA, and HH, each lasting 10 s.
Figure 5.
Raw fNIRS channels (measuring changes in ) selected after the proposed channel selection algorithm featuring the relative range (RR). The intervals for various pain conditions are highlighted and annotated as B (Baseline), LA (Low Arm Pain), HA (High Arm Pain), HH (High Hand Pain), and LH (Low Hand Pain). The gray background in the figure represents the duration of each experiment phase: Baseline: 60 s, LA, LH, HA, and HH, each lasting 10 s.
Figure 5.
Raw fNIRS channels (measuring changes in ) selected after the proposed channel selection algorithm featuring the relative range (RR). The intervals for various pain conditions are highlighted and annotated as B (Baseline), LA (Low Arm Pain), HA (High Arm Pain), HH (High Hand Pain), and LH (Low Hand Pain). The gray background in the figure represents the duration of each experiment phase: Baseline: 60 s, LA, LH, HA, and HH, each lasting 10 s.
Figure 6.
Raw 10-Second Data Segments for Baseline (B), Low Pain (LP), and High Pain (HP) Classes, displayed for Channel 1 of (Left) and (Right).
Figure 6.
Raw 10-Second Data Segments for Baseline (B), Low Pain (LP), and High Pain (HP) Classes, displayed for Channel 1 of (Left) and (Right).
Figure 7.
Preprocessed 10-Second Data Segments for Baseline (B), Low Pain (LP), and High Pain (HP) Classes, displayed for (Left) and (Right). The processing pipeline encompasses low-pass filtering, Common Average Referencing (CAR) for each filtered channel, and the final step of averaging across all channels, culminating in a consolidated vector representation.
Figure 7.
Preprocessed 10-Second Data Segments for Baseline (B), Low Pain (LP), and High Pain (HP) Classes, displayed for (Left) and (Right). The processing pipeline encompasses low-pass filtering, Common Average Referencing (CAR) for each filtered channel, and the final step of averaging across all channels, culminating in a consolidated vector representation.
Figure 8.
Haemodynamic changes shown using fNIRS for (first row) and (second row) measures: (a) Baseline, (b) HH (High Hand Pain), (c) LH (Low Hand Pain), (d) HA (High Arm Pain), and (e) LA (Low Arm Pain). The color bar signifies the change in concentration of and (µmol). These calculations are derived from the averages across all subjects for each respective channel.
Figure 8.
Haemodynamic changes shown using fNIRS for (first row) and (second row) measures: (a) Baseline, (b) HH (High Hand Pain), (c) LH (Low Hand Pain), (d) HA (High Arm Pain), and (e) LA (Low Arm Pain). The color bar signifies the change in concentration of and (µmol). These calculations are derived from the averages across all subjects for each respective channel.
Figure 9.
Class-wise accuracy (%) assessment of different measures using Disc; KNN; and SVM classifiers using confusion charts.
Figure 9.
Class-wise accuracy (%) assessment of different measures using Disc; KNN; and SVM classifiers using confusion charts.
Table 1.
Details of statistical features used in this study. The feature vector F comprises all ten features, with h as the preprocessed signal ( or ), as the derivative of h, and as the mean of . , , and denote the peak, root mean square, and absolute mean of the input signal h, respectively, while represents the variance.
Table 1.
Details of statistical features used in this study. The feature vector F comprises all ten features, with h as the preprocessed signal ( or ), as the derivative of h, and as the mean of . , , and denote the peak, root mean square, and absolute mean of the input signal h, respectively, while represents the variance.
Features | Definitions |
---|
Log Energy | |
Crest Factor | |
Shape Factor | |
Impulse Factor | |
Margin Factor | |
Mobility | |
Complexity | |
Mean Absolute Deviation of First Derivative | |
Range | |
Variation in First Derivative | |
Table 2.
Optimised hyperparameters for different classification algorithms via Bayesian Optimisation in the context of distinguishing between Baseline (B), Low Pain (LP), and High Pain (HP).
Table 2.
Optimised hyperparameters for different classification algorithms via Bayesian Optimisation in the context of distinguishing between Baseline (B), Low Pain (LP), and High Pain (HP).
Model | Parameters | | | + |
---|
Disc | Discriminant Type | Pseudo Linear | Linear | Diagonal Linear |
Gamma | 7.55 × 10−4 | 0.0025 | 0.006 |
Delta | 3.51 × 10−5 | 2.96 × 10−5 | 2.12 × 10−5 |
KNN | Number of Neighbours | 211 | 1 | 25 |
Distance | Chebychev | Cosine | City Block |
Distance Weight | Inverse | Inverse | Equal |
Exponent | – | – | – |
Neighbour Search | KD-Tree | Exhaustive | Exhaustive |
Standardisation | Yes | Yes | Yes |
SVM | Coding | One vs. All | One vs. All | One vs. One |
Box Constraint | 2.1888 | 10.3923 | 980.4894 |
Kernel Scale | – | – | 13.2018 |
Kernel Function | Polynomial | Polynomial | Gaussian |
Polynomial Order | 3 | 3 | – |
Standardise | Yes | Yes | Yes |
Table 3.
System performance metrics (Acc: Accuracy, Sen: Sensitivity, Spec: Specificity, and F1 Score) for different classification algorithms (Disc, KNN, and SVM) across various measures, with each measure having a different feature vector length denoted by #.
Table 3.
System performance metrics (Acc: Accuracy, Sen: Sensitivity, Spec: Specificity, and F1 Score) for different classification algorithms (Disc, KNN, and SVM) across various measures, with each measure having a different feature vector length denoted by #.
Measure | Model | # | Acc | Sen | Spec | F1 Score |
---|
| Disc | 10 | | | | |
KNN | | | | |
SVM | | | | |
| Disc | 10 | | | | |
KNN | | | | |
SVM | | | | |
| Disc | 20 | | | | |
KNN | | | | |
SVM | | | | |
Table 4.
System performance metrics (Acc: Accuracy, Sen: Sensitivity, Spec: Specificity, and F1 Score) with MRMR-based selected features for different classification algorithms (Disc, KNN, and SVM) applied to each measure, with the feature vector length denoted by #.
Table 4.
System performance metrics (Acc: Accuracy, Sen: Sensitivity, Spec: Specificity, and F1 Score) with MRMR-based selected features for different classification algorithms (Disc, KNN, and SVM) applied to each measure, with the feature vector length denoted by #.
Measure | Model | # | Acc | Sen | Spec | F1 Score |
---|
| Disc | 10 | 51.78 ± 9.94 | 74.78 ± 19.43 | 73.30 ± 11.91 | 85.98 ± 9.60 |
KNN | 7 | 44.22 ± 8.16 | 55.36 ± 15.38 | 70.22 ± 13.16 | 76.30 ± 7.20 |
SVM | 9 | 65.71 ± 5.97 | 93.18 ± 8.03 | 95.99 ± 4.24 | 96.77 ± 3.67 |
| Disc | 10 | 50.94 ± 7.6 | 73.57 ± 12.12 | 75.77 ± 9.53 | 85.35 ± 6.42 |
KNN | 10 | 41.83 ± 8.34 | 44.36 ± 14.3 | 74.23 ± 9.10 | 73.14 ± 5.75 |
SVM | 9 | 63.42 ± 6.85 | 94.44 ± 8.33 | 97.22 ± 3.27 | 97.40 ± 3.84 |
| Disc | 20 | 56.23 ± 6.84 | 76.32 ± 11.62 | 79.32 ± 10.81 | 87.24 ± 5.73 |
KNN | 18 | 40.8 ± 7.26 | 44.58 ± 15.27 | 68.83 ± 9.34 | 71.72 ± 6.34 |
SVM | 15 | 68.51 ± 9.02 | 94.70 ± 5.77 | 94.29 ± 4.92 | 97.33 ± 2.92 |
Table 5.
List of selected features for each measure, with # indicating the number of features.
Table 5.
List of selected features for each measure, with # indicating the number of features.
Measure | # | Selected Features |
---|
| 9 | Mobility, Complexity, Range, Shape Factor, Variation in First Derivative, Impulse Factor, Mean Absolute Deviation of First Derivative, Log Energy, Crest Factor. |
| 9 | Crest Factor, Complexity, Shape Factor, Mobility, Range, Variation in First Derivative, Log Energy, Mean Absolute Deviation of First Derivative, Margin Factor. |
| 15 | : Mobility, Complexity, Range, Shape Factor, Variation in First Derivative, Impulse Factor, Mean Absolute Deviation of First Derivative, Log Energy, Crest Factor. : Crest Factor, Complexity, Shape Factor, Mobility, Range, Variation in First Derivative. |
Table 6.
Post Hoc Test Results for Different Levels of Pain in Various Features of (Only comparisons with significant () values are reported.)
Table 6.
Post Hoc Test Results for Different Levels of Pain in Various Features of (Only comparisons with significant () values are reported.)
Feature | Group One | Group Two | Mean Diff. | Std. Error | Sig. | Lower Bound | Upper Bound |
---|
Log Energy | No Pain | Low Pain | 73.67 | 41.110 | 0.073 | −7.00 | 154.33 |
High Pain | 96.99 * | 41.110 | 0.018 | 16.33 | 177.66 |
Low Pain | No Pain | −73.67 | 41.110 | 0.073 | −154.33 | 7.00 |
High Pain | 23.33 | 40.966 | 0.569 | −57.06 | 103.71 |
High Pain | No Pain | −96.99 * | 41.110 | 0.018 | −177.66 | −16.33 |
Low Pain | −23.33 | 40.966 | 0.569 | −103.71 | 57.06 |
Crest factor | No Pain | Low Pain | −0.02 | 0.039 | 0.685 | −0.09 | 0.06 |
High Pain | 0.078 * | 0.039 | 0.044 | 0.00 | 0.15 |
Low Pain | No Pain | 0.02 | 0.039 | 0.685 | −0.06 | 0.09 |
High Pain | 0.094 * | 0.039 | 0.015 | 0.02 | 0.17 |
High Pain | No Pain | −0.078 * | 0.039 | 0.044 | −0.15 | 0.00 |
Low Pain | −0.094 * | 0.039 | 0.015 | −0.17 | −0.02 |
Shape factor | No Pain | Low Pain | −0.017 * | 0.006 | 0.008 | −0.03 | 0.00 |
High Pain | −0.01 | 0.006 | 0.067 | −0.02 | 0.00 |
Low Pain | No Pain | 0.017 * | 0.006 | 0.008 | 0.00 | 0.03 |
High Pain | 0.01 | 0.006 | 0.402 | −0.01 | 0.02 |
High Pain | No Pain | 0.01 | 0.006 | 0.067 | 0.00 | 0.02 |
Low Pain | −0.01 | 0.006 | 0.402 | −0.02 | 0.01 |
Impulse factor | No Pain | Low Pain | −0.05 | 0.055 | 0.386 | −0.16 | 0.06 |
High Pain | 0.08 | 0.055 | 0.167 | −0.03 | 0.19 |
Low Pain | No Pain | 0.05 | 0.055 | 0.386 | −0.06 | 0.16 |
High Pain | 0.125 * | 0.055 | 0.024 | 0.02 | 0.23 |
High Pain | No Pain | −0.08 | 0.055 | 0.167 | −0.19 | 0.03 |
Low Pain | −0.125 * | 0.055 | 0.024 | −0.23 | −0.02 |
Range | No Pain | Low Pain | −0.165 * | 0.037 | | −0.24 | −0.09 |
High Pain | −0.129 * | 0.037 | 0.001 | −0.20 | −0.06 |
Low Pain | No Pain | 0.165 * | 0.037 | | 0.09 | 0.24 |
High Pain | 0.04 | 0.037 | 0.33 | −0.04 | 0.11 |
High Pain | No Pain | 0.129 * | 0.037 | 0.001 | 0.06 | 0.20 |
Low Pain | −0.04 | 0.037 | 0.33 | −0.11 | 0.04 |
Table 7.
Post Hoc Test Results for Different Levels of Pain in Various Features of (Only comparisons with significant p-values are reported).
Table 7.
Post Hoc Test Results for Different Levels of Pain in Various Features of (Only comparisons with significant p-values are reported).
Feature | Group One | Group Two | Mean diff. | Std. Error | Sig. | Lower Bound | Upper Bound |
---|
Log Energy | No Pain | Low Pain | 104.153 * | 49.629 | 0.036 | 6.770 | 201.535 |
High Pain | 110.774 * | 49.629 | 0.026 | 13.391 | 208.156 |
Low Pain | No Pain | −104.153 * | 49.629 | 0.036 | −201.535 | −6.770 |
High Pain | 6.62 | 49.4554 | 0.894 | −90.420 | 103.662 |
High Pain | No Pain | −110.774 * | 49.629 | 0.026 | −208.156 | −13.391 |
Low Pain | −6.62 | 49.455 | 0.894 | −103.662 | 90.420 |
Margin Factor | No Pain | Low Pain | 1.629 * | 0.572 | 0.004 | 0.506 | 2.752 |
High Pain | 0.621 | 0.572 | 0.277 | −0.501 | 1.744 |
Low Pain | No Pain | −1.629 * | 0.572 | 0.004 | −2.752 | −0.506 |
High Pain | −1.007 | 0.570 | 0.078 | −2.126 | 0.111 |
High Pain | No Pain | −0.621 | 0.572 | 0.277 | −1.744 | 0.501 |
Low Pain | 1.007 | 0.570 | 0.078 | −0.111 | 2.126 |
Range | No Pain | Low Pain | −0.106 * | 0.04 | 0.00 | −0.18 | −0.04 |
High Pain | −0.072 * | 0.04 | 0.04 | −0.14 | 0.00 |
Low Pain | No Pain | 0.106 * | 0.04 | 0.00 | 0.04 | 0.18 |
High Pain | 0.03 | 0.04 | 0.34 | −0.04 | 0.10 |
High Pain | No Pain | 0.072 * | 0.04 | 0.04 | 0.00 | 0.14 |
Low Pain | −0.03 | 0.04 | 0.34 | −0.10 | 0.04 |