PaperlayerView exact revision on arXiv

Source-linked AI summary

A Multimodal Autonomic Sensing Framework for Objective Assessment of Patient Responses to Dental Pulp Stimulation

Youngsun Kong, Yubin Choi, Dongjin Song, Dong-Guk Shin, I-Ping Chen, Ki Chon

arXiv:2609.35121v1cs.LGeess.SP

TL;DR

Dental pain is subjective and can be difficult to communicate, motivating objective assessment of responses to pulp stimulation. The study combines autonomic signals in a multimodal framework and reports stronger binary than three-class classification performance.

  • Problem

    Dental pain is subjective and difficult to communicate, and objective quantitative measures are lacking for assessment.

    Source: §I · p. 1

  • Method

    The framework integrates RRI, EDA, and SKNA using attention-based mid-level fusion to assess pulp-stimulation responses.

    Source: §I · p. 2

  • Results

    80.2% balanced accuracy for binary classification and 60.0% balanced accuracy for multiclass classification were achieved.

    Source: §IV · p. 8

  • Takeaways & Limitations

    The findings support the feasibility of non-invasive autonomic monitoring as an adjunct tool for pulp vitality assessment.

    Source: §V · p. 10

    The intermediate class showed reduced separability from adjacent classes, and autonomic outputs reflect global responses influenced by confounding factors.

    Source: §V · p. 11

Abstract

from arXiv · show

Patient responses to dental pulp testing, ranging from no sensation to intense pain, provide important information for assessing pulp status in endodontic diagnosis. However, pain is a subjective sensory and emotional experience that varies considerably across individuals and can be difficult to communicate. We investigated whether complementary autonomic signals could support objective assessment of responses during dental examination. Forty-nine patients underwent cold pulp testing, yielding no-response, mild-response, and intense-response conditions. The framework integrated ECG-derived skin nerve activity (SKNA) and R-R intervals (RRI), together with electrodermal activity (EDA), using temporal convolutional network encoders with attention-based mid-level fusion. Individual baseline signals and subject-level covariates, including anxiety scores and biological sex, were also incorporated. The framework achieved 80.2% balanced accuracy, 75.2% sensitivity, and 85.2% specificity for binary classification of no response versus mild or intense response. For three-class classification, it achieved 60.0% balanced accuracy and a 58.8% macro-averaged F1 score. Ablation and attention-weight analyses indicated that EDA contributed most strongly to model performance, followed by RRI, while SKNA improved balanced accuracy by approximately five percentage points. Age was significantly associated with model performance. These findings support the feasibility of multimodal autonomic sensing for objective, non-invasive assessment of responses to dental pulp stimulation.

I. INTRODUCTION

The introduction motivates objective, physiology-based assessment because dental pain is subjective and difficult to communicate, while existing cardiovascular measures have limitations for capturing rapid responses. It proposes combining complementary autonomic signals to assess responses to pulp stimulation.

I. INTRODUCTION

  • The introduction places the clinical need in the context of caries-related pulp infection, which can lead to hypersensitivity, severe pain, and rare life-threatening complications.It states that caries is highly prevalent and its burden continues to rise with population growth.Source: §I · p. 1
  • Subjective pain ratings and the lack of objective quantitative measures complicate dental-pain assessment, including for patients with limited communication abilities.The introduction notes potential challenges of misdiagnosis and inappropriate analgesic prescribing.Source: §I · p. 1
  • HR and HRV require cautious interpretation because they do not directly index sympathetic activity and can miss rapid, transient responses in acute testing.These measures are influenced by parasympathetic regulation and are often analyzed over relatively long windows.Source: §I · pp. 1–2
  • The proposed framework combines RRI, EDA, and SKNA with attention-based mid-level fusion to assess responses to pulp stimulation.RRI and SKNA can be derived from a surface ECG setup, while EDA and SKNA reflect distinct sympathetic pathways.Source: §I · p. 2

A. Automatic Assessment of Dental Pain and Pulp Responses

Prior work has explored automatic dental-response detection using multimodal behavioral and physiological signals, but relatively little work has characterized distinct response patterns during pulp testing. The present study addresses that gap with autonomic sensing during cold stimulation.

A. Automatic Assessment of Dental Pain and Pulp Responses

  • An early EPT system combined muscle activity, finger movement, and vocal responses, stopping stimulation when any modality detected a pulp-evoked response.The approach used predefined thresholds rather than machine learning or learned multimodal representations.Source: A. Automatic Assessment of Dental Pain… · p. 2
  • The review describes a relative paucity of automatic pulp-stimulation studies compared with broader automatic pain-recognition research.The BioVid Heat Pain Database, introduced in 2013, provided experimentally induced heat-pain responses and physiological and facial recordings.Source: A. Automatic Assessment of Dental Pain… · p. 2
  • Teichmann et al. used ECG, PPG, and respiratory effort from 20 patients for general dental pain-event detection, while an EDA study focused on clinically meaningful pain during EPT.The reported random-forest sensitivity and specificity were 87% and 63%; the EDA study reported 76% sensitivity and 87% specificity for VAS-defined pain.Source: A. Automatic Assessment of Dental Pain… · p. 2
  • Facial-expression analysis has also been explored, but clinician, hand, or instrument occlusion may make facial sensing challenging during pulp testing.Source: A. Automatic Assessment of Dental Pain… · p. 2
  • Automatic pulp-response research has largely detected response presence or pain intensity, leaving limited investigation of clinically defined patterns from no response through mild sensation to intense pain.The present study targets these response categories during cold pulp stimulation.Source: A. Automatic Assessment of Dental Pain… · p. 2

B. Automatic Pain Recognition Using Machine Learning

Automatic pain recognition has developed from handcrafted-feature classifiers toward deep temporal and attention-based multimodal fusion, often evaluated on public pain datasets. Building on these approaches, this study targets clinically relevant response categories during dental pulp stimulation.

B. Automatic Pain Recognition Using Machine Learning

  • Public pain-recognition datasets combine physiological signals such as ECG, PPG, EMG, respiration, and EDA with behavioral indicators, and multimodal combinations have often outperformed single modalities.The review names BioVid, SenseEmotion, X-ITE, PainMonit, and AI4PAIN among established datasets.Source: B. Automatic Pain Recognition Using… · pp. 2–3
  • Earlier BioVid studies used handcrafted features and conventional classifiers, while later CNN- and RNN-based approaches learned temporal or modality-specific representations.Reported accuracies include 80.6% and 77.8% for personalized and generalized RF models, 79% for a CNN approach, and 83.3% for a hybrid RNN–ANN approach.Source: B. Automatic Pain Recognition Using… · p. 3
  • Recent methods use attention or Transformers to weight modality features or model cross-modal dependencies, with reported BioVid accuracies from 82.7% to 87.5%.The cited examples include attention-enhanced physiological fusion, facial-video and heart-rate integration, and separate EDA and ECG modeling followed by inter-modal learning.Source: B. Automatic Pain Recognition Using… · p. 3
  • The proposed framework adapts modality-specific CNN encoders and attention-based fusion to dental testing, accounting for clinical sensing constraints and responses beyond pain intensity alone.Its target categories include no response, mild sensation, and intense pain; facial sensing may be limited by occlusion in dental settings.Source: B. Automatic Pain Recognition Using… · p. 3

A. Participants

Participants were recruited from endodontic clinics among patients requiring root canal treatment for one tooth and having two nearby healthy teeth. The final cohort comprised 49 adults after one enrolled participant’s data were excluded for wire disconnection.

A. Participants

  • Eligible participants required root canal treatment for one tooth and had two nearby healthy teeth.Source: A. Participants · p. 3
  • Exclusions included minors, pregnancy, porcelain crowns, medications with anticholinergic side effects, prior sympathectomy, or Raynaud’s syndrome.The medication criterion was intended to address effects on skin conductance.Source: A. Participants · p. 3
  • 50 participants enrolled, and data from one were excluded because of wire disconnection, leaving 49 participants aged 21–68 years.The final cohort included 21 males and 28 females.Source: A. Participants · p. 3

B. Dental stimulus — cold testing

Cold pulp stimulation was followed by patient VAS ratings and dentist classification into no, mild, or intense responses. The protocol also included sham trials and recovery intervals, and targeted two response-classification tasks.

B. Dental stimulus — cold testing

  • Responses were clinically adjudicated as no response, mild response (+), or intense response (++), alongside a 0–10 VAS rating after each stimulus.The dentist applied refrigerant to a cotton pellet placed briefly on the target tooth; participants raised a hand when they perceived the stimulus.Source: B. Dental stimulus — cold testing · p. 3
  • Fig. 1 presents VAS score distributions for negative, mild (+), and intense (++) cold-testing outcomes.The figure caption defines the three outcome labels but does not state their score distributions.Source: B. Dental stimulus — cold testing · p. 3, Figure 1 · p. 4
  • The protocol tested at least three teeth per participant, included a clean-cotton sham, and used interstimulus intervals of at least 20 seconds.At least one affected tooth and two nearby healthy teeth were tested.Source: B. Dental stimulus — cold testing · p. 4
  • The two tasks distinguish negative from positive responses to assess vitality, and negative/mild from intense responses to identify heightened responses associated with pulpitis.Depending on pulp status, vital or inflamed teeth may produce mild-to-intense sensations, whereas necrotic teeth typically produce no response.Source: B. Dental stimulus — cold testing · p. 4

C. Data Acquisition

EDA and ECG were acquired with clinical-grade devices using electrodes on the non-dominant hand and lower limbs. Participants rested seated before cold stimulation.

C. Data Acquisition

  • EDA was recorded from the index and middle fingers of the non-dominant hand, while ECG electrodes were placed on the inner wrist and ankles.The reference electrode was positioned on the opposite ankle; the non-dominant hand left the dominant hand available to signal stimulus onset.Source: C. Data Acquisition · p. 4
  • Signals were sampled at either 10 kHz or 4 kHz using a PowerLab amplifier and LabChart 8 software.EDA and ECG were acquired using a clinical-grade GSR Amp device and BioAmp system, respectively.Source: C. Data Acquisition · p. 4
  • Before electrode placement, skin was cleaned with 70% isopropyl alcohol, and participants rested seated for at least two minutes before stimulation.Source: C. Data Acquisition · p. 4

D. Feature extraction

The feature pipeline derives time-resolved inputs from EDA and RRI and windowed descriptive and complexity features from SKNA. Sex and dental anxiety scores are included as auxiliary covariates.

Electrodermal activity (EDA)

  • EDA was decomposed into phasic and tonic components, and 11 features—including phasic derivatives, phasic driver, and spectrogram information—were extracted after resampling and filtering.The phasic component captures rapid stimulus-evoked changes, whereas the tonic component reflects slower baseline sympathetic tone.Source: D. Feature extraction · p. 4

R-R intervals (RRI)

  • RRI and its first-order derivative were used instead of conventional HRV indices, which typically require data lengths on the order of one minute.R-peaks were detected from resampled ECG, manually reviewed and corrected, and the resulting RRI was resampled to 4 Hz.Source: D. Feature extraction · pp. 4–5

Skin nerve activity (SKNA)

  • SKNA features were computed in 5-second windows with 4.5-second overlap, producing time-resolved sequences at 2 Hz to balance model complexity given limited data.Eleven features included amplitude, variability, Hjorth, fractal-dimension, and entropy measures.Source: D. Feature extraction · p. 5
  • SKNA was isolated from downsampled, high-pass-filtered ECG, with participant-specific narrowband artifacts identified and removed before TVSKNA computation.TVSKNA was derived by VFCDM decomposition into 12 bands, reconstruction over 480–1120 Hz, Hilbert-transform amplitude estimation, and 100 ms moving-average smoothing.Source: D. Feature extraction · p. 5

Subject-level covariates

  • Biological sex and dental anxiety scores were selected as auxiliary covariates based on cohort distribution and their reported influence on autonomic responses.The four-item DAS totals 4–20, with moderate anxiety at 9–12, high at 13–14, and severe at 15–20.Source: D. Feature extraction · p. 5

E. Proposed Framework

The framework processes EDA, RRI, and SKNA with modality-specific encoders, then combines their representations through attention-based fusion. It also incorporates baseline signals and subject-level covariates, with evaluation designed to hold out participants.

Proposed Framework

  • Each modality has an independent encoder, and the framework integrates EDA, RRI, and SKNA representations using attention-based fusion alongside sex and anxiety covariates.Domain-specific time-series features are encoded before fusion; the attention output is concatenated with auxiliary features for classification.Source: E. Proposed Framework · pp. 5, 7, Figure 3 · p. 5
  • The temporal convolutional module captures modality-specific patterns, while EDA and RRI use TCN encoders and shorter SKNA sequences use a residual CNN.Adaptive average pooling reduces sequences into four temporal bins; each encoder produces a 64-dimensional embedding.Source: E. Proposed Framework · pp. 6–7
  • Baseline adjustment emphasizes each signal’s relative deviation from baseline rather than its absolute magnitude.A learnable parameter limits the influence of each modality’s deviation from baseline.Source: E. Proposed Framework · p. 7
  • Auxiliary features modulate modality representations through scale and shift before fusion, with the design intended to keep covariates from dominating physiological signals.The scaling and shifting factors are empirically set to 0.5 and 0.01, respectively.Source: E. Proposed Framework · p. 7
  • Evaluation uses leave-one-subject-out cross-validation, with training and validation splits kept subject-independent and stratified by class.The validation set supports early stopping and learning-rate reduction.Source: E. Proposed Framework · p. 7
  • The study compares the proposed model with CNN- and TCN-based concatenation baselines and conventional classifiers, using shared time-series features for the conventional models.Interpretation also examines modality attention, baseline-adjustment coefficients, auxiliary-feature sensitivity, and associations between performance and demographic variables.Source: E. Proposed Framework · pp. 7–8

IV. RESULTS

The framework outperformed conventional late-fusion tabular baselines, with performance and modality contributions varying across classification tasks.

IV. RESULTS

  • 80.2% binary and 60.0% three-class balanced accuracy were achieved by the framework; the corresponding F1 scores were 78.5% and 58.8%.Table II summarizes performance for negative versus positive responses and for negative, mild, and intense responses.Source: §IV · p. 8, Table II · p. 9
  • 72.3% and 52.8% were the best conventional baseline's binary and multiclass balanced accuracies; CNN-Concat and TCN-Concat each reached 79.5% binary balanced accuracy.For multiclass classification, CNN-Concat and TCN-Concat achieved 55.1% and 57.2% balanced accuracy, respectively.Source: §IV · p. 9
  • 80.2% to 66.5% binary balanced accuracy followed EDA removal; three-class accuracy fell to 51.8% and 51.7% after EDA and RRI removal, respectively.Source: §IV · p. 9
  • Removing baseline and auxiliary features shifted binary balanced accuracy from 80.2% to 79.1% and 79.9%, and three-class accuracy from 60.0% to 59.9% and 58.7%, respectively.The authors characterize these components as complementary with minimal but additive contributions.Source: §IV · p. 9
  • Mild responses were more often confused with adjacent classes, while the model reasonably distinguished stimulation from negative responses and separated negative from intense responses.The authors attribute mild-response errors as likely reflecting autonomic-response variability, anxiety, individual cold sensitivity, and inter-subject variability.Source: §IV · p. 9
  • Attention weights assigned EDA the largest binary contribution; in multiclass classification, EDA and RRI were comparably important and exceeded SKNA.Source: §IV · p. 9
  • RRI relied more on baseline adjustment than other modalities and was also more influenced by auxiliary features.The authors relate this to RRI's biphasic dynamics and prior findings on HRV modulation by biological sex and anxiety.Source: §IV · p. 9
  • For binary classification, age was significantly associated with performance (OR = 1.03, 95% CI [1.01, 1.07], p = 0.031); the multiclass association was not significant.The multiclass estimate was OR = 1.02, 95% CI [1.00, 1.05], p = 0.09.Source: §IV · p. 9, Table IV · p. 10

V. DISCUSSION

The multimodal model distinguished dental stimulation from negative responses, with EDA contributing most strongly and RRI next; SKNA also improved balanced accuracy. Age was associated with performance, while intermediate-class separability, dataset size, and signal artifacts constrain interpretation and deployment.

V. DISCUSSION

  • For multi-class classification, race showed a non-significant trend toward higher performance in white subjects (OR = 1.84, 95% CI [0.97, 3.77], p = 0.07).The authors note the training dataset included 34 white and 15 non-white participants, which may contribute to the trend.Source: §V · p. 10
  • EDA was the most influential modality, followed by RRI, while SKNA contributed approximately a 5% improvement in balanced accuracy.The authors note that SKNA performance may vary by task and that clinical noise, motion, and anxiety-related confounders may affect its reliability.Source: §V · p. 11
  • Age emerged as an important variable influencing model performance, and the proposed explanation involving reduced baseline variability requires further validation.The authors suggest reduced baseline variability may increase contrast between resting and stimulus-induced responses.Source: §V · p. 11
  • Intermediate-class separability was reduced, and the limited dataset may be affected by racial imbalance in the multi-class results.The authors suggest more direct pulp-specific or central nervous system measurements and larger, more diverse datasets as directions for addressing these constraints.Source: §V · p. 11
  • Data were collected under relatively controlled conditions, and real-world deployment will require automated artifact detection, particularly given limited preliminary work for SKNA.The authors identify robust artifact detection and removal as future development needs.Source: §V · p. 11
  • The framework demonstrated reliable performance distinguishing dental stimulation from negative responses, supporting its potential as an objective tool for pulp vitality assessment.The authors describe the overall approach as feasible for objective detection of dental pain.Source: §V · pp. 10–11

V. CONCLUSION

The conclusion presents a noninvasive multimodal framework using EDA, SKNA, and RRI that performed strongly for negative-versus-positive classification and reasonably for multi-class classification. Dataset size, class separability, and signal artifacts remain constraints on real-world applicability.

V. CONCLUSION

  • The framework used noninvasive EDA, SKNA, and RRI signals and showed strong negative-versus-positive classification performance, reasonable multi-class performance, and better results than conventional tabular feature-based approaches.Demographic analysis suggested age may influence performance, while other factors showed limited effects.Source: §V · p. 11
  • Dataset size, class separability, and signal artifacts remain limitations, motivating broader datasets, additional modalities, and robust artifact detection for real-world applicability.The authors frame noninvasive physiological monitoring as a potential objective and scalable tool for pulp vitality assessment in clinical settings.Source: §V · pp. 11–12
Loading 2609.35121v1…