Few-Shot Learning for Personalised Automated Pain Assessment

Heinke Hihn; Ibrahim Eisawy; Patrick Thiam; Hans A. Kestler; Friedhelm Schwenker

cs.LGOct 7, 2026

Few-Shot Learning for Personalised Automated Pain Assessment

Authors: Heinke Hihn, Ibrahim Eisawy, Patrick Thiam, Hans A. Kestler, Friedhelm Schwenker

Organizations: IU International University of Applied Sciences, Berlin, Germany · Institute of Neural Information Processing, Ulm University, Ulm, Germany · Institute of Medical Systems Biology, Ulm University, Ulm, Germany

Abstract

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.

Figures & tables

Fig 1 : The data in the experiments has been generated by subjecting subjects to various levels of pain (annotated as

T_{i}

) and a baseline of no pain and recording their bodily response by various bio-physical signals such as EDA, EMG, and ECG. This creates time series of signals annotated by the experienced pain level, thus enabling us to train models to predict the pain based on the signals.

Panel of Figure 2 — Fig 2 : The left figure shows the signal segmentation in the SenseEmotion Database [ 56 ] , where experiments have been carried out on $6.5s$ windows with a $4s$ shift from the onset of the elicitation in four pain levels $T_{0},...,T_{3}$ . On the right, we show how in the BioVid Database [ 57 ] the windows are of $4.5s$ with an $4s$ shift on five pain levels $T_{0},...,T_{4}$ . Thus, on sampling rate of 256 Hz we have an input dimensionality $1,664$ of and $1,152$ , respectively.

Panel of Figure 3 — Fig 3 : Our approach for Subject-As-A-Task Few-Shot Pain Assessment. First, we sample a task, which is then passed into an encoder for each modality. These streams are combined using a modified Cross Modality Transformer as introduced by Farmani et al. [ 12 ] , as described in Section 3.3 . The resulting support feature maps are used to compute a prototype per class against which we classify the query feature maps using the Cross Alignment Network described in Section 3.4 .

Panel of Figure 4 — Fig 4 : Our method supports two modes of inference: in the Zero-Shot setting (left), we use a global precomputed bank to initialize the prototypes. Thus, this mode does not require any labelled samples from the subject. In the $k$ -Shot case (right), we use a portion of the held-out subjects samples to initialize the prototypes (the support set) and classify the remainder (the query set). Thus, the Zero-Shot setup is the aggregated population level classifier and the $k$ -Shot setup is the personalised individual classifier.

Method	Binary	Four Classes
Thiam et al. [ 51 ]	83.39%	43.89%
Thiam et al. [ 49 ]	81.05%	40.80%
Thiam et al. [ 54 ]	81.50%	n/a
Thiam et al. [ 53 ]	64.35%	n/a
Kessler et al. [ 32 ]	71.85%	n/a
Ours (Zero-Shot)	82.37% ( $\pm 10.9$ )	41.88% ( $\pm 6.6$ )

Table 1 : Overview of reported pain recognition performance by algorithm/model on the SenseEmotion dataset. All values are reported for Leave-One-Subject-Out (LOSO) cross-validation. Accuracy is reported with standard deviation in parentheses where available. Bold indicates best Zero-Shot method and underlined indicates that the personalised

k

-Shot classifiers beats previous methods.

Method	Binary	Five Classes
Farmani et al. [ 12 ]	87.52% ( $\pm 11.0$ )	n/a
Aslam et al. [ 4 ]	86.90%	n/a
Li et al. [ 35 ]	86.21%	38.03%
Lu et al. [ 39 ]	85.56%	34.46%
Thiam et al. [ 54 ]	85.32% ( $\pm 13.7$ )	n/a
Jiang et al. [ 30 ]	84.58% ( $\pm 13.3$ )	39.24% ( $\pm 8.65$ )

Table 2 : Overview of reported pain recognition performance by algorithm/model on the BioVid dataset. All values are reported for Leave-One-Subject-Out (LOSO) cross-validation. Accuracy is reported with standard deviation in parentheses where available. Bold indicates best Zero-Shot method and underlined indicates that the personalised

k

-Shot classifiers beats previous methods.

Method	Binary
Luebke et al. [ 40 ]	93.62%
Gouverneur et al. [ 17 ]	93.26%
Gouverneur et al. [ 17 ]	89.79% ( $\pm$ 11.68)
Gouverneur et al. [ 16 ]	87.41% ( $\pm 11.99$ )
Gouverneur et al. [ 17 ]	91.09% ( $\pm 9.70$ )
Ours (Zero-Shot)	90.47% ( $\pm$ 9.23)

Table 3 : Overview of reported pain recognition performance by algorithm/model on the PainMonit dataset. All values are reported for Leave-One-Subject-Out (LOSO) cross-validation. Accuracy is reported with standard deviation in parentheses where available. Bold indicates best Zero-Shot method and underlined indicates that the personalised

k

-Shot classifiers beats previous methods.

	BioVid
Method	T 0 vs. T 1	T 0 vs. T 2
Thiam et al. [ 48 ]	61.15% ( $\pm 12.20$ )	66.81% ( $\pm 15.92$ )
Gkikas et al. [ 13 ]	52.38%	52.78%
Werner et al. [ 63 ]	48.70%	51.60%
Gkikas et al. [ 13 ]	62.82%	63.68%
Lopez & Picard [ 38 ]	56.44%	59.40%

Table 4 : Comparison of additional binary pain recognition results. Bold indicates best Zero-Shot method and underlined indicates that the personalised

k

-Shot classifiers beats previous methods.

Fig 5 : Personalisation benefits per subject. A vast majority of 91% to 98% of subjects benefit from personalisation in the BioVid full classification setting, i.e., using part of their samples to initialize the prototypes (see Panels A, B, D). Panel A shows a mean improvement in macro F1 is $+0.097$ . Panel E shows the distribution of the improvements in the different $k$ settings.

Fig 6 : Personalisation benefits per class on the BioVid dataset. Panels A and B show that the intermediate pain levels 1, 2, and 3 have the largest mean F1 gains from personalisation. Their respective F1 scores improve on average, indicating that personalisation contributes most for pain levels that vary only little in magnitude, compared with the extremes 0 and 4. Panels C and D show how subjects benefit on a per-class basis, indicating the same pattern of improvement in classes 1 to 3.

Fig 7 : Different settings of training for the BioVid (A+B) and the SenseEmotion (C+D) dataset. In all settings increasing the training task size (

k,q

values) improves the results with a ceiling on

k,q=10

where we do not observe significantly higher results. All significant

p

-values are

<0.001

Fig 8 : Mean held-out performance across training and evaluation conditions for binary classification (top) and multi-class classification (bottom) on the SenseEmotion dataset. Columns show accuracy and macro-F1. The Zero-Shot condition $(k_{e},q_{e})=(0,20)$ uses learned prototype memory, whereas the remaining conditions use target-subject support.

Fig 9 : Effect of training and evaluation task size on BioVid binary (A+B) and five-class (C+D) performance. Points represent held-out subjects, and boxes show the score distributions. Here, T and E denote the training and configured

k

-shot evaluation tasks, respectively.

	BioVid		SenseEmotion
Modalities	Zero-Shot	$k$ -Shot	Zero-Shot	$k$ -Shot
EMG+RSP	N/A	N/A	63.79% $(\pm 11.6)$	64.71% $(\pm 11.7)$
ECG+RSP	N/A	N/A	64.60% $(\pm 11.3)$	67.51% $(\pm 12.2)$
EDA+RSP	N/A	N/A	81.64% $(\pm 10.4)$	82.44% $(\pm 10.7)$
EDA+EMG	85.52% $(\pm 9.31)$	85.68% $(\pm 9.31)$	81.52% $(\pm 9.31)$	81.99% $(\pm 10.7)$
ECG+EMG	59.39% $(\pm 15.2)$	65.14% $(\pm 16.0)$	58.46% $(\pm 10.1)$	63.34% $(\pm 11.8)$

Table 5: Ablation study on the modalities used. In accordance with previous studies [ 39 , 35 ] , EDA+ECG achieves the highest classification accuracy in the binary (

T_{0}

vs.

T_{4}

and

T_{0}

vs.

T_{3}

, respectively) LOSO setting. N/A indicates that the corresponding modality is not available in the dataset. We report accuracy and trained the models with

k,q=10

Stage	Core configuration	Output
Input	EDA and ECG windows with 1152 steps	$B\times 1152\times 2$
Modality encoders	Independent temporal convolutional frontends with pooling	$2\times(B\times 78\times 16)$
CrossMod fusion	Two 8-head self-attention layers followed by bidirectional cross-modal attention	$B\times 78\times 32$
Prototype memory	$M_{\mathrm{slot}}=2$ learned temporal prototype slots per class	$B_{t}\times\|\mathcal{C}\|M_{\mathrm{slot}}\times 78\times 32$
Query encoding	The same modality encoders and CrossMod fusion are applied to each query	$B_{t}\times Q\times 78\times 32$
CAN alignment	Pair-conditioned bidirectional temporal attention and attended cosine similarity, $\tau=0.025$	$B_{t}\times Q\times\|\mathcal{C}\|M_{\mathrm{slot}}$

Table 6 : Compact architecture of the proposed CAN model.

Hyperparameter	Setting
Temporal frontend	Filters $8/16$ ; kernels $64/16$
Pooling / dropout / $L_{2}$	$4,8$ / $0.25$ / $10^{-4}$
CrossMod configuration	2 layers; 8 heads; hidden dim. 128
CAN configuration	$\tau=0.025$ ; meta hidden dim. 32
CAN support source	Learned prototype memory
Prototype memory	2 slots/class; 256 init. samples/class

Table 7 : Main hyperparameters used for the full LOSO experiment.

Explore similar work

Sep 27, 2026cs.CV

Prompt-Anchored Residual Adaptation for Biomedical Vision-Language Models

Pretrained biomedical vision-language models achieve strong zero-shot performance in biomedical image classification. However, downstream biomedical classification often depends on subtle visual differences between classes that may not be fully captured by pretrained representations. Few-shot adaptation addresses this mismatch by optimizing a task-specific predictor on a small labeled support set. Because the selected examples capture only part of the visual variation within the target classes, the adapted predictions can depend strongly on their composition. We propose Prompt-Anchored Residual Adaptation (PARA), which retains the frozen prompt prediction as a support-invariant semantic anchor and incorporates a visual prediction learned from the support set through an anchor-relative residual. The residual step is computed in a closed form from frozen support embeddings using anchor discrepancy and support agreement. Support-set dependence also limits evaluation: comparisons are fair within a shared draw but remain conditional on its composition. To obtain more reliable comparisons, we introduce a repeated-support protocol that separates support-selection variation from optimization randomness and reports both average and worst-20% performance. PARA achieves state-of-the-art performance in both few-shot classification and base-to-novel generalization.

Jingxuan Kang, Qianying Yue, Che Liu +1

Imperial College London · The Chinese University of Hong Kong

Few-Shot Learning Vision-Language Model Adaptation Additive-Residual Selective Invariance

Ask PDF

May 6, 2026cs.CV

Few-Shot Learning Pipeline for Monkeypox Skin Disease Classification Using CNN Feature Extractors

Despite the strong performance of Convolutional Neural Networks (CNNs) in disease classification, their effectiveness often depends on access to large annotated datasets, which is an impractical requirement for emerging or rare conditions such as Monkeypox. To overcome this limitation, we propose a few-shot learning (FSL) framework that employs SimpleShot, a lightweight, non-parametric, inductive classifier, for Monkeypox and pox-like skin disease recognition from limited labeled examples. The proposed pipeline passes the skin lesion images through a frozen, pretrained CNN backbone to obtain feature embeddings, which are then classified via SimpleShot using nearest-centroid comparisons in a normalized embedding space. We systematically benchmark six widely used CNN backbones as feature extractors under consistent experimental settings, enabling fair comparison. Experiments on three publicly available datasets (MSLD v1.0, MSID, and MSLD v2.0) are conducted across 2-way, 4-way, and 6-way tasks with 1-shot, 5-shot, and 10-shot configurations. Among all models, MobileNetV2_100 consistently achieves the highest accuracy. In addition, we present a cross-dataset evaluation for Monkeypox classification, revealing that binary Mpox-vs-Others transfer remains comparatively stable while multi-class performance degrades significantly under domain shift. Together, these results demonstrate the practical utility of combining inductive FSL methods with lightweight CNN backbones and highlight the importance of domain robustness for reliable real-world clinical deployment.

Md. Safirur Rashid, Sabbir Ahmed, Muhammad Usama Islam +2

Department of Computer Science and Engineering, Islamic University of Technology, Gazipur, Bangladesh · Management Information Systems, Metropolitan State University, Minnesota, USA · School of Kinesiology, University of Louisiana at Lafayette, LA, USA

Few-Shot Learning Skin Lesion Classification Convolutional Neural Networks

Ask PDF

Date pendingcs.CV

Are We Really Doing Few-Shot Learning? A Critical Examination of Pre-Training Assumptions

Few-shot learning is commonly evaluated under protocols that pre-train a model on a large auxiliary set whose classes are disjoint from the target episodes yet drawn from the same visual domain. This paper examines whether such protocols truly reflect low-data learning. We systematically compare no pre-training, class-disjoint in-domain pre-training, supervised out-of-domain pre-training, and label-free out-of-domain pre-training across eight datasets, three few-shot architectures, and multiple way-shot settings. Our results show that class disjointness alone is insufficient to remove the influence of target-domain data. In-domain pre-training improves over no pre-training by 33.41 percentage points on average, whereas supervised out-of-domain pre-training yields 23.75 percentage points, revealing a 9.66-point optimistic bias associated with domain overlap. Although out-of-domain pre-training is more realistic in applications where target-domain data are scarce, its effectiveness depends strongly on the compatibility between source and target domains. We further show that labeled source data are not strictly required, with an augmentation-based label-free strategy reaching an average gain of 27.71 percentage points and closely matching supervised out-of-domain pre-training at 27.97 percentage points. Finally, we introduce a descriptor-based source-selection strategy that estimates source-domain suitability before pre-training, reaching a median gap of only 1.37 percentage points to oracle selection. These findings highlight the need to move beyond in-domain pre-training as the default few-shot evaluation protocol, since it can overestimate performance in realistic scenarios where target-domain data are scarce.

Alejandro Galan-Cuenca, Marcelo Saval-Calvo, Antonio Javier Gallego

University Institute for Computer Research, University of Alicante, Alicante, Spain

Few-Shot Learning Target Domain Pretraining Unlabeled Data Limited Data

Ask PDF