Pain is a subjective experience that can be expressed by either verbal or non-verbal behaviors [25]. Body movements, vocalizations and facial expressions are some examples of the non-verbal expressions of pain [26]. Several scales have been designed to assess the level of pain. The Critical Care Pain Observation Tool (CPOT), the Pain Assessment in Advanced Dementia Scale (PAINAD) and FLACC scale are objective measures of pain in which facial expressions are taking into consideration while assessing the pain level [27]. It was shown that facial expressions are one of the basic structures in pain detection [27]. Pain communication through facial expression is well established and validated [28][29].
Recently, advances in technology, followed by developing many analytic means in the area of artificial intelligence (AI) help researchers in progressing automatic detection of facial expressions of pain [30]. This study aimed to develop an artificial intelligence (AI) software for pain assessment in children while receiving an IANB and to evaluate its performance compared to human examiners who used FLACC scale.
The studied AI software (Achydtct) was trained to use the Child Facial Coding System (CFCS) to evaluate facial expressions in preschool children by detecting the ten involved Action Units (AUs). CFCS is a valid and reliable system that is used to assess the intensity of children pain [31, 32]. As facial expressions related to pain can vary between different races, we used the results concluded by the study of Alkhouli et al, [9] which was performed on Syrian children to determine the AUs associated with each level of pain. FLACC scale is a valid and reliable scale that is used commonly in the field of pediatric dentistry in assessing pain level. To ensure the accuracy of the scoring process, two experienced raters were used to analyze the videos. After that, inter- and intra-rater reliability tests were performed. a high level of agreement between the raters and within each one was detected.
The age group of 6–9 years was selected to be involved in this study because children in this age group participate in the same emotional and motor development, which makes their facial expressions similar and close [33]. Each child received an inferior nerve block injection. This was because the inferior alveolar nerve block is a commonly used local anesthesia technique in dental procedures involving the mandibular molars and the associated soft and hard tissues [34].
The choice of the Inferior Alveolar Nerve Block (IANB) injection for performing pulpotomy on primary mandibular second molars is underpinned by several considerations. Firstly, the IANB provides profound anesthesia to the mandibular molars, ensuring effective pain management during the pulpotomy procedure. Additionally, the IANB allows for a broader and more predictable coverage of the affected area, enhancing the success and efficiency of the pulpotomy. The age group for children incorporates into this study is more suitable for block than infiltration to anesthetize primary mandibular 2nd molars [35].
Pain level was detected by the two measures (FLACC and Achydtct) at a specific time point of the procedure's pain triggers. We selected needle insertion as a specific point of pain reference, which allows for the comparison of pain levels between different patients or procedures. This can help to identify potential factors that may contribute to variations in pain levels and inform the development of interventions to reduce pain during dental procedures.
The AI model used in this study is a convolutional neural network (CNN). CNNs are a type of deep learning model commonly used in image recognition and analysis tasks [36]. In the current study, the CNN was trained to analyze images of children undergoing dental procedures and predict their pain level based on the CFCS.
In the realm of AI applications within clinical settings, ethical considerations play a pivotal role in safeguarding patient rights and privacy. The deployment of the AI technology investigated in this study raises pertinent ethical concerns that warrant careful deliberation. Specifically, the issue of patient consent emerges, as individuals may question their autonomy regarding the use of AI in dental procedures. It is imperative to address whether patients should have the right to refuse the application of this AI technology by dentists. Furthermore, the potential misuse of patient images collected by the AI system poses a critical ethical dilemma, necessitating a robust discussion on data security measures and the prevention of unauthorized use [37, 38].
The use of CNNs in pain assessment is a relatively new field of research, but previous studies have shown promising results [39]. CNNs can be trained to identify patterns and features in images that may not be visible to human raters, potentially improving the accuracy and reliability of pain assessment [39]. Furthermore, CNNs can process large amounts of data quickly, making them a potentially useful tool in busy clinical settings [40].
In the present study, the CNN was trained on a dataset of about 4000 images of children undergoing dental procedures. The dataset was divided into four levels of pain that range from 0: no pain to 3: severe pain. Two trained researchers with CFCS conducted the classification of the images and high level of agreement was resulted.
Our findings showed that Achydtct has high level of sensitivity, specificity and accuracy (0.87, 0.92 and 0.90, respectively). This means that the developed software can correctly identify children with\without dental injections causing pain. Our findings are in line with previous studies that have validated the use of AI in pain assessment. In a study by Xu et al (2018), an AI-powered system achieved an accuracy of 80% in detecting pain in children after appendectomy procedures [41]. However, our AI software showed higher level of accuracy and this can be attributed to the difference in the pain measure used between the two studies. In the study of Xu et al, they used a 0–10 numerical rating scale (self-report measure of pain) which is considered less accurate than the FLACC scale in children. Similarly, Wu et al. (2022) reported a higher accuracy rate in detecting children pain from their facial expressions using an AI software analysis system than in assessing the pain by human volunteers [42]. These results suggest that AI-based pain assessment systems have the potential to accurately and reliably detect pain in pediatric patients.
The Kappa coefficient showed that there was an almost perfect agreement between the values of pain level detected by the AI software and by FLACC in each grade of pain intensity. However, the highest kappa value was shown in detecting grade 3 (severe pain). This can be attributed to the increased contraction of facial expression muscles when severe pain is felt, which makes it easier for the AI software to detect the Action units.
The present study found no significant difference in pain assessment between the AI software and the FLACC scale, neither in boys nor in girls. This result can be attributed to the high sensitivity and specificity of the AI software, which has been validated in this study.
Moreover, the lack of difference in pain assessment between boys and girls is consistent with previous studies that have reported no significant sex differences in pain perception among children [43, 44]. This suggests that Achydtct is equally effective in assessing pain in both boys and girls, regardless of potential differences in pain perception.
This paper introduces a novel approach in the field by investigating the use of artificial intelligence to assess pain intensity during local dental anesthetic injections in pediatric patients aged 6–9 years. The innovative aspect lies in the application of machine learning algorithms to analyze facial action units associated with pain, providing an objective and potentially more accurate measure compared to traditional subjective assessments.
The limitation of this study should be mentioned. The study was conducted in a single center, single city, which may not reflect the complexity and variability of pain assessment in different races and cultures. For that reason, further research is needed to validate the Achydtct performance in other settings and with diverse patient populations.