Background. Although telehealth strategies can be effectively used for managing cancer pain, identifying the best care pathway for tailoring interventions and allocating resources remains difficult. Artificial intelligence and machine learning (ML) may help clinicians develop more accurate strategies for predicting whether patients need remote consultations or in-person evaluations. Methods. Data from two cohorts of cancer pain patients were analyzed. Variables included sociodemographic and clinical data, including age, sex, ECOG performance status, metastases, bone metastases, pain type, breakthrough cancer pain (BTCP), and rapid onset opioids (ROOs) therapy. The main outcome was the number of televisits (one versus multiple). After harmonizing the dataset, categorical variables were one-hot encoded, and age was standardized. Six models were tested: logistic regression, random forest (RF), gradient boosting machine (GBM), support vector machine (SVM), k-nearest neighbors (KNN), and multilayer perceptron (MLP). Training and tuning used a 7-repeated 5-fold cross-validation approach. Performance was evaluated on a hold-out test set using F1-score, accuracy, and AUC-ROC. A sensitivity analysis with two scenarios was performed to verify the effects of class weighting and excluding the cohort variable. Results. The final dataset included 270 patients. No variable was significantly linked to the number of televisits. F1-scores across models ranged from 0.33 (RF) to 0.65 (MLP), accuracy from 0.45 (RF) to 0.55 (SVM), and AUC-ROC from 0.43 (RF) to 0.65 (LR). DeLong tests showed no significant differences between algorithms (p > 0.05). While the MLP achieved the highest F1-score, it showed instability with 91% null F1-scores. Incorporating class weights slightly improved SVM (F1 = 0.58; AUC = 0.62) and LR (F1 = 0.53; AUC = 0.63), though not significantly. Removing the cohort variable reduced training time by about one hour and yielded similar results. Conclusion. Although no model demonstrated strong predictive power, this ML-based framework shows the potential of using structured telemedicine data to model clinical workload and optimize follow-up strategies in cancer pain care. Further studies with feature-rich datasets are needed to improve clinical usefulness.

Background: Although telehealth strategies can be effectively adopted to manage cancer pain, identifying the optimal care pathway for tailoring interventions and allocating resources remains difficult. Artificial intelligence and machine learning (ML) may help clinicians develop more accurate strategies for predicting whether patients need remote consultations or in‐person evaluations. Methods: Data from two cohorts of cancer pain patients were analyzed. Variables included sociodemographic and clinical data, including age, sex, ECOG performance status, metastases, bone metastases, pain type, breakthrough cancer pain (BTCP), and rapid onset opioids (ROOs) therapy. The main outcome was the number of televisits (one versus multiple). For preprocessing, datasets from the two cohorts were harmonized by aligning variable definitions, coding schemes, and data formats. Six models were tested: logistic regression, random forest (RF), gradient boosting machine (GBM), support vector machine (SVM), k‐nearest neighbors (KNNs), and multilayer perceptron (MLP). Training and tuning used a 7‐repeated 5‐fold cross‐validation approach. Performance was evaluated on a hold‐out test set using F 1‐score, accuracy, and AUC‐ROC. A sensitivity analysis with two scenarios was performed to verify the effects of class weighting and excluding the cohort variable. Results: The final dataset included 270 patients. No statistically significant associations were identified between the available variables and the number of televisits. F 1‐scores across models ranged from 0.33 (RF) to 0.65 (MLP), accuracy from 0.45 (RF) to 0.55 (SVM), and AUC‐ROC from 0.43 (RF) to 0.65 (LR). DeLong tests showed no significant differences between algorithms ( p > 0.05). Although the MLP achieved the highest F 1‐score, it exhibited instability, with 91% of null F 1‐scores. Incorporating class weights slightly improved SVM ( F 1 = 0.58 and AUC = 0.62) and LR ( F 1 = 0.53 and AUC = 0.63) though not significantly. Conclusion: Although no model demonstrated strong predictive power, this ML‐based framework shows the potential of using structured telemedicine data to model clinical workload and optimize follow‐up strategies in cancer pain care. Trial Registration: ClinicalTrials.gov identifier: NCT04726228 and NCT07038434

Although emotions play a fundamental role in modulating pain perception, their objective assessment in clinical contexts remains challenging. Recent advances in artificial intelligence (AI) have opened new opportunities to measure emotional states through facial expression analysis, physiological signal modeling, natural language processing (NLP), and multimodal data integration. In affective computing, the field that focuses on technologies designed to recognize, interpret, process, and simulate human emotions, facial expression-based emotion recognition has progressed from traditional machine learning methods to advanced deep learning approaches, including convolutional neural networks (CNNs), attention-based hybrid models, and transformer architectures. Similarly, recurrent neural networks and self-supervised learning methods have been implemented for developing models from physiological signals such as electrocardiography, photoplethysmography, galvanic skin response, and related biosignals. Additionally, NLP systems can extract affective information from naturalistic text, using both lexicon-based and transformer-based models. Finally, multimodal fusion and alignment techniques allow the integration of heterogeneous data streams, providing richer and more ecologically valid emotion representations. Collectively, these strategies offer powerful tools for advancing automatic pain assessment (APA) in cancer care, with the potential to support personalized, emotion-aware therapeutic approaches. However, from an AI perspective, several open challenges remain, including multimodal representation learning under weak supervision, robustness to missing or degraded modalities, limited explainability of affective inference models, lack of standardized benchmarking protocols, and the presence of bias and domain shift in emotion datasets. Given the inherently subjective, context-dependent, and culturally mediated features of the emotional experience, further research is needed to address these technical limitations, integrating technological advances with the intrinsic complexity of emotion interpretation.