BACKGROUND The objective recording of subjectively experienced pain is a problem that has not been sufficiently solved to date. In recent years, data sets have been created to train artificial intelligence algorithms to recognize patterns of pain intensity. The multimodal recognition of pain with machine learning could provide a way to reduce an over- or undersupply of analgesics, explicitly in patients with limited communication skills. OBJECTIVES This study investigated the methodology of automated multimodal recognition of pain intensity and modality using machine-learning techniques of artificial intelligence. Multimodal recognition rates of experimentally induced phasic electrical and heat pain stimuli were compared with uni- and bimodal recognition rates. MATERIAL AND METHODS On the basis of the X‑ITE Pain Database, healthy subjects were stimulated with phasic electro-induced pain and heat pain, and their corresponding pain responses were recorded with multimodal sensors (acoustic, video-based, physiological). After complex signal processing, machine-learning methods were used to calculate recognition rates with respect to pain intensity (baseline vs. pain threshold, pain tolerance, mean value of pain threshold and tolerance) and pain modality (electrical vs. heat). Finally, a statistical comparison of uni- vs. multimodal and bi- vs. multimodal detection rates was performed. RESULTS With few exceptions, multimodal recognition of pain intensity rates was statistically superior to unimodal recognition rates, regardless of the pain modality. Multimodal pain recognition distinguished significantly better between heat and electro-induced pain. Further, multimodal recognition rates were predominantly superior to bimodal recognition rates. CONCLUSION Priority should be given to the multimodal approach to the recognition of pain intensity and modality compared with unimodality. Further clinical studies should clarify whether multimodal automated recognition of pain intensity and modality is in fact superior to bimodal recognition.

中文翻译：

---

Multimodale Erkennung von Schmerzintensität und -modalität mit maschinellen Lernverfahren

背景技术主观体验疼痛的客观记录是迄今为止尚未得到充分解决的问题。近年来，已经创建了数据集来训练人工智能算法来识别疼痛强度的模式。通过机器学习对疼痛进行多模态识别可以提供一种减少镇痛药供应过剩或供应不足的方法，尤其是在沟通技巧有限的患者中。目标本研究调查了使用人工智能机器学习技术自动多模式识别疼痛强度和模式的方法。将实验诱导的相位电和热痛刺激的多模态识别率与单模态和双模态识别率进行比较。材料和方法 基于 X‑ITE 疼痛数据库，健康受试者受到阶段性电诱导疼痛和热痛的刺激，并用多模式传感器（声学、视频、生理）记录他们相应的疼痛反应。经过复杂的信号处理后，机器学习方法用于计算疼痛强度（基线与疼痛阈值、疼痛耐受性、疼痛阈值和耐受性的平均值）和疼痛方式（电与热）的识别率。最后，对单模式与多模式以及双模式与多模式检测率进行了统计比较。结果 除了少数例外，疼痛强度率的多模态识别在统计学上优于单模态识别率，无论疼痛模式如何。多模式疼痛识别在热痛和电诱导痛之间的区别明显更好。更多，多模态识别率主要优于双模态识别率。结论 与单模态相比，应优先考虑多模态方法来识别疼痛强度和模态。进一步的临床研究应该澄清疼痛强度和方式的多模态自动识别实际上是否优于双模态识别。