Source-linked AI summary
Few-Shot Learning for Personalised Automated Pain Assessment
Heinke Hihn, Ibrahim Eisawy, Patrick Thiam, Hans A. Kestler, Friedhelm Schwenker
TL;DR
Pain’s subjectivity creates substantial inter-subject variability in population-trained classifiers, motivating models that generalize across individuals. The paper evaluates few-shot personalized classifiers, conditioning on support samples from the held-out subject to classify query samples under leave-one-subject-out evaluation. In LOSO evaluation, k-Shot uses a subset of the held-out subject’s data to calibrate the model.
Problem
Pain’s subjectivity creates substantial inter-subject variability in population-trained classifiers, motivating models that generalize across individuals.
Source: §1 · p. 2Method
The paper evaluates few-shot personalized classifiers, conditioning on support samples from the held-out subject to classify query samples under leave-one-subject-out evaluation.
Source: §3.2 · pp. 8–9Results
In LOSO evaluation, k-Shot uses a subset of the held-out subject’s data to calibrate the model.
Source: §3.2 · p. 9Takeaways & Limitations
The k-Shot evaluation scheme calibrates the model using a subset of held-out-subject data.
Source: §3.2 · p. 9K-shot personalization requires labelled target-subject examples for every class, so annotation costs grow as N × K and may be substantial in clinical settings.
Source: §5.1 · p. 21
Abstract
from arXiv · showhide
Pain perception varies substantially across individuals, making it difficult for population-based classifiers to generalise across all subjects in a dataset. One way to account for subject variability is to train personalised classifiers. In this work, we evaluate Few-Shot Learning, a sub-area of Meta-Learning, as an approach to personalisation in automated pain assessment. We re-interpret the shift from population-level to subject-level evaluation as a task-domain shift, where the observed classes remain fixed but the target subject changes. We evaluate our method on the BioVid Pain Database, the SenseEmotion Database, and the PainMonit Experimental Dataset (PMED), reaching 85.75% and 35.49% accuracy on BioVid and 82.37% and 41.88% on SenseEmotion in the binary and multi-class settings under a Leave-One-Subject-Out CV protocol respectively, and 90.47% on PMED, for which only a binary benchmark exists. Using samples to implement k-shot conditioning, the accuracies can be improved to 86.25%, 40.06%, 83.43%, 44.08%, and 91.25%, respectively. To further evaluate the effects and robustness of our method, we provide additional ablation experiments and investigate the personalisation effects. Our results suggest that support-conditioned few-shot adaptation can improve average performance under inter-subject variability.
Author summary · 1 Introduction
Pain is subjective and shaped by personal factors, so the same stimulus can produce different responses and population-trained assessment models face substantial inter-subject variability. The introduction motivates Few-Shot Learning as a way to adapt pain-assessment models to individuals from limited labelled data.
1 Introduction
- Pain is subjective and shaped by cognition, past experience, and current circumstances, causing the same stimulus to produce different reactions or none at all.Source: §1 · pp. 1–2
- Because patients may be unable to self-report, caregivers use visible bodily reactions and available biophysical signals; automated systems could support physicians’ pain-level decisions.Self-reporting tools such as the Visual Analogue Scale (VAS) are the most common method.Source: §1 · p. 2
- Controlled pain-stimulus studies record EDA, ECG, EMG, and facial video, enabling assessment methods from handcrafted features to deep learning.Source: §1 · p. 2
- Inter-subject variability challenges population-trained classifiers; responses span preprocessing, model-level, population-plus-individual, and generalisation-focused adaptation approaches.Examples include subject-aware normalisation, personalised models, multi-task and mixed-effects modelling, and Few-Shot-style adaptation.Source: §1 · p. 2
- Few studies have evaluated Meta-Learning for automated pain assessment, motivating its investigation as a way to reduce inter-subject variability.Few-Shot Learning adapts to a new task or individual from a small number of labelled examples, leveraging prior knowledge rather than learning from scratch.Source: §1 · p. 2
- Pain-assessment experiments typically provide only 8 to 20 samples per class and subject, making Few-Shot Learning suited to the combination of limited labels and subject variability.Source: §1 · pp. 2–3
- The introduction frames its central research question as how much personalisation can improve automated pain assessment under inter-individual variability.Source: §1 · p. 3
2. Do the benefits of personalisation vary across individuals? · 3. Do the benefits of personalisation vary across pain levels?
The paper presents a few-shot approach to personalised pain assessment, with a prototype-scoring network and a zero-shot extension.
3. Do the benefits of personalisation vary across pain levels?
- Few-Shot Learning is proposed for building personalised classifiers.This is presented as an approach to address the paper’s research questions about personalisation.Source: §3 · p. 3
- The Cross Alignment Network efficiently scores samples against class-prototypes.Source: §3 · p. 3
- A Zero-Shot extension evaluates samples without requiring labelled support samples.Source: §3 · p. 3
2 Related Work
Related work connects few-shot learning with cross-subject evaluation, prototype-based pain assessment, and methods for personalising pain models to individual subjects or patients.
2.1 Combining Few-Shot Learning with LOSO Evaluation
- Few-shot LOSO uses labelled support samples from a held-out subject for adaptation or calibration, unlike strict LOSO, and pain studies include Rathee et al.’s meta-learning approach.Rathee et al. report higher accuracy for 1-shot and 5-shot learning than their baselines; other pain applications target limited labelled data and inter-individual variability.Source: §2.1 · pp. 3–4
2.2 Prototype-Based Classification in Pain Assessment
- Prototype-based pain methods span fMRI pain decoding, chronic-pain patient clustering, and transfer from phasic to tonic stimuli.Ricken et al. report improved classification across most distance metrics, including TRA electric pain accuracy rising from 66.0% to 68.0% and thermal EDA from 53.2% to 59.6%.Source: §2.2 · pp. 4–5
2.3 Personalisation in Automated Pain Assessment
- Personalisation approaches include neutral-video adaptation with generated target expressions, sensitivity-aware feature fusion, and transfer-component adaptation for new patients.DSFDA was evaluated on BioVid and UNBC-McMaster; Jiang et al. evaluated on SpaExp and BioVid; ATLAS used BioVid ECG, skin conductance, and EMG.Source: §2.3 · p. 5
- 65.7% to 66.8% and 67.6% to 67.8% were the modest gains Thiam et al. reported on left- and right-forearm datasets over a baseline using all data.The authors attributed modest overall accuracy to participants who showed little visible facial reaction to the stimuli.Source: §2.3 · pp. 5–6
3 Method
The method frames subject-level pain assessment as task-based few-shot learning: support examples form class prototypes for query classification, and multimodal EDA/ECG features are fused. It evaluates this approach across BioVid, SenseEmotion, and PMED datasets.
3.1.1 The BioVid Pain Dataset
- BioVid includes 87 participants and four thermal-pain levels; the dataset has 8,700 samples, expanded to 87,000 by augmentation, and experiments use EDA and ECG.Each elicitation lasted 4 seconds; pain levels were interspersed with T0 recovery periods.Source: §3.1.1 · pp. 6–7
3.1.2 The SenseEmotion Dataset
- SenseEmotion includes 45 participants exposed to three pain levels, yielding 4,050 pain-stimulus samples; experiments use EDA and ECG.Each stimulus lasted 4 seconds, with randomized 8–13-second pauses between stimuli.Source: §3.1.2 · p. 7
3.1.3 The PainMonit Dataset
- PMED includes 52 participants exposed to five thermal-pain levels, with eight 10-second elicitation repetitions per level; experiments use EDA and ECG.The recorded signals also include EMG, BVP, respiration, temperature, and grip.Source: §3.1.3 · p. 7
3.2 Few-Shot Task Construction
- In LOSO evaluation, the held-out subject is excluded from training tasks; support samples form class prototypes for query classification, with Zero-Shot and k-Shot conditioning schemes.Zero-Shot uses a prototype bank built from training subjects, whereas k-Shot calibrates on a subset of the held-out subject’s data.Source: §3.2 · pp. 8–9
3.3 CrossMod Feature Fusion
- The multimodal approach uses a modified CrossMod feature-map encoder to combine EDA and ECG signal streams, based on Farmani et al..The encoder operates on signal windows of length T.Source: §3.3 · p. 9
F , xECG
The method encodes EDA and ECG separately before cross-modal fusion, then scores query maps against class prototypes with temporal cross-attention. A multi-slot prototype bank enables zero-shot inference without subject labels, while support-based prototypes enable k-shot inference.
F , xECG
- Separate EDA and ECG encoders feed prenormalised Transformer stacks; bidirectional cross-attention then fuses their maps into one temporal representation, with no separate classifier.The same encoder produces support maps, which are averaged class-wise into prototypes for subsequent scoring.Source: F , xECG · pp. 9–10
3.4 Cross-Attention Network Scoring
- CAN scores each prototype–query pair using a descriptor of their temporal means, difference, and element-wise product, then combines attention-weighted summaries with unweighted means before cosine similarity.Attention is based on a temporal cosine-similarity correlation matrix; a learned shared gate controls the interpolation between weighted summaries and means.Source: §3.4 · pp. 10–11
3.5 Zero-Shot Classification via a Prototype Bank
- A multi-slot prototype bank substitutes for support prototypes in zero-shot classification, requiring no labelled samples from the held-out subject.Each slot averages encoded samples drawn without replacement across training subjects: 258 samples per class and slot, or three per each of 86 training subjects; k-shot inference initializes prototypes from held-out support samples.Source: §3.5 · p. 11, Figure 4 · p. 12
4 Results
Across BioVid, SenseEmotion, and PMED, support-conditioned k-shot adaptation raised accuracy over zero-shot evaluation; personalisation benefits varied across subjects and pain levels, and ablations examined task setup and signal channels.
4.5 Experimental Setup
- Experiments used a single 16 GB consumer-grade GPU on Google Colab; evaluation averaged metrics on held-out subjects in one LOSO run, using augmented training data and original, unaltered evaluation samples.Hyperparameters were selected by an Optuna search with a reduced number of iterations.Source: §4 · p. 12, §4.5 · p. 18
4.1 Personalisation for Automated Pain Assessment
- 85.75% and 35.49% BioVid, 82.37% and 41.88% SenseEmotion, and 90.47% PMED zero-shot accuracies rose to 86.25%, 40.06%, 83.43%, 44.08%, and 91.25% k-shot, respectively.The accuracy pairs are binary and multi-class for BioVid and SenseEmotion; PMED has only a binary benchmark. K-shot uses held-out-subject labels and is not directly comparable to zero-shot methods. K-shot multi-class accuracy exceeded state-of-the-art results on BioVid (40.06% vs. 39.24%) and SenseEmotion (44.08% vs. 43.89%).Source: §4.1 · pp. 12–13, Table 1 · p. 13, Table 2 · p. 13
- In additional binary settings, k-shot helped across classification problems, with Farmani et al. the exception on BioVid, reporting 87.52% versus our 86.25%.Zero-shot performance was higher for T0 vs. T1 and T0 vs. T2 on BioVid, and T0 vs. T2 on SenseEmotion; T0 vs. T1 was the most difficult setting.Source: §4.1 · p. 14, Table 4 · p. 15
4.3 Personalisation Benefits per Class
- Mean F1 gains were 0.190, 0.183, and 0.174 for classes 1–3, versus −0.126 for class 0 and 0.061 for class 4.Personalisation redistributed performance toward intermediate pain levels, reducing class-0 F1 and, at small k, F1 for the highest pain class.Source: §4.3 · p. 15, Figure 6 · p. 17
4.2 Personalisation Benefits per Subject
- In BioVid 10-shot evaluation, 91%–98% of subjects benefited, with mean macro-F1 gain 0.097, but individual changes ranged from −0.03 to 0.33.At k = 1, noisy or atypical support samples could produce unrepresentative prototypes; five or more samples increased the fraction benefiting. The 500 support/query episodes estimate expected episodic performance, so unfavorable samples could lower a subject’s estimate.Source: §4.2 · pp. 14–15, Figure 5 · p. 16
4.4.1 Training and Evaluation Task Setup
- During training, scores improved significantly up to kt, qt = 5, with smaller gains beyond that point; the effect hits a ceiling after 10 samples.During evaluation, more support samples improved average accuracy only in personalised k-shot settings, enabling less noisy, more representative prototypes.Source: §4.4.1 · p. 16
- Binary zero-shot accuracy and macro-F1 were unaffected by training configuration, while k-shot metrics varied significantly; five-class zero-shot accuracy stayed stable, but k-shot accuracy varied strongly.For five-class tasks, macro-F1 varied in both phases, with a larger effect for k-shot than zero-shot performance.Source: §4.4.1 · pp. 16–17
4.4.2 Choice of Signal Channels
- EDA and ECG together yielded the highest performance among tested two-modality pairs, but the difference was marginal and not statistically significant.The ablations suggest the Few-Shot Subject-as-a-Task setup learns robust, transferable feature maps; further tests should vary k while holding q fixed.Source: §4.4.2 · pp. 17–18
5 Discussion
Support-conditioned k-Shot adaptation was associated with higher mean performance than Zero-Shot, especially in multi-class tasks, but benefits varied across subjects and pain classes. The findings are limited by the evaluation protocol and experimental setting, motivating fixed-query comparisons and clinical validation.
5 Discussion
- 40.06% on BioVid and 44.08% on SenseEmotion were k-Shot multi-class accuracies, numerically above listed comparisons but not evidence of superiority given labelled target support and study differences.K-Shot had higher mean accuracy than Zero-Shot in every evaluated setting; the results indicate performance attainable when a small, balanced set of subject-specific labels is available.Source: §5 · pp. 18–19, Figure 8 · p. 19
- Most held-out subjects had higher estimated macro-F1 with k-Shot, but some did not benefit, and differing query sets prevent interpreting the differences as fixed-query effects.Several support samples may yield a more stable class prototype than one atypical or noisy sample, but this mechanism was not directly measured.Source: §5 · pp. 19, 21, Figure 9 · p. 20
- Intermediate BioVid pain levels showed the largest F1 gains, while class-0 F1 decreased and maximum-pain-class gains were smaller, indicating error redistribution rather than uniform improvement.Greater variability or overlap for intermediate stimuli is a possible explanation, but the experiments did not test it.Source: §5 · p. 21
- The prototype-based ECG-and-EDA approach treats each held-out subject as a new task and adapts without updating all model parameters, but requires labelled examples for every target class.This complements prior personalised pain-assessment and Few-Shot methods while requiring more target-subject labels than approaches using unlabelled data or a neutral reference alone.Source: §5 · p. 21
5.1 Limitations
- The evaluation uses transductive Zero-Shot normalisation, sampled episodes, and tenfold augmentation, while omitting seed-variation analysis and a common comparison pipeline.The datasets involve healthy participants, controlled heat stimuli, and ECG and EDA; clinical validation is needed before generalising to clinical pain or less controlled settings.Source: §5.1 · pp. 21–22
5.2 Future Work
- Future work should use paired fixed-query evaluations and repeated training seeds, compare prototype and normalisation alternatives, and develop sequential adaptation with fewer balanced-label requirements.It should also evaluate calibration, macro-F1, and clinically relevant error costs alongside accuracy.Source: §5.2 · p. 22
5.3 Conclusion
- Under the present episodic protocol, subject-specific support was associated with higher mean performance, particularly for multi-class tasks and intermediate pain levels, but not uniformly across subjects or classes.Zero-Shot needs no labelled target samples but currently assumes access to an unlabelled subject batch; fixed-query comparisons and clinical validation remain necessary.Source: §5.3 · p. 23