The Challenge of Assessing Pain: Can Artificial Intelligence Objectively Measure a Subjective Experience?
Welcome to Health Connect, the podcast for health professionals through which we will share the latest news and information on science and technology in the medical field. Today, we will talk about two machine learning models for pain assessment.
Did you know that the International Association for the Study of Pain has recently revised its definition of pain? Its current description refers to pain as "an unpleasant sensory and emotional experience associated with, or resembling that associated with, actual or potential tissue damage". Several contextualizing notes were added, acknowledging that pain is always a personal experience that is influenced by personal, psychological, and social factors, as well as the importance of respecting the report of the pain experience. Finally, it is emphasized that verbal description is only one of many possible behaviors with which pain is expressed.1
Often considered "the fifth vital sign", pain is an essential symptom or sign of numerous medical conditions. In clinical settings, it is generally assessed through patient self-report,2 usually with verbal reporting techniques. These assessments are often considered the gold standard, but due to their subjective nature, they may not be reliable.1,2
One of the main limitations of verbal assessments is the uncertainty of how they relate to the underlying experience of the patient. It is important to note that a lack of concordance has been reported between patients using the same analog scales, and even low consistency was found when individual patients actively interpret the meaning of their experiences. Likewise, biases of self-presentation can distort the report; for example, one of the most common is the stoicism bias, where enduring pain is associated with perceived masculinity. This bias could partially explain why, in both clinical and experimental studies, men usually report less pain than women.1
Another important deficiency of the verbal report is found in the instances in which it is impossible to obtain it. For example, when the patient is unable to describe their pain, or cannot do so reliably. This can occur with infants, people with verbal communication problems, certain types of dementia, and critically ill patients.1 In intensive care settings, pain assessment based on facial expressions is a critical tool.3
To discuss advances in pain detection through facial observation, let us first establish the two general categories of pain indicators, physiological and behavioral. The first category, physiological indicators, includes autonomic responses and activations in the central nervous system. The second category, behavioral indicators, includes both verbal descriptions and facial expressions, along with acts of withdrawal or avoidance, called "instrumentals," and vocalizations or grimaces, known as "expressive acts".1
Since its initial studies, pain indicators emphasized the role of observation and the need for precision when defining and evaluating behaviors, parameters that are shared with the so-called "study of emotion". This area of study was driven in part by the work of Charles Darwin. In The Expression of the Emotions in Man and Animals, Darwin argued that the affective states, including pain, could be shared phylogenetically with other species and are represented in specific behavioral topographies, especially in the facial expression.1
Continuing the history of the study of facial movements, in the late 1970s, Ekman and Friesen published the Facial Action Coding System. This is an anatomically based, atheoretical, and relatively objective system that deconstructs any facial movement into "action units", based on the changes that appear when a muscle or combination of muscles is activated. This system is composed of 44 action units, and most can be described in terms of their intensity.1
The Facial Action Coding System, as well as similar adapted systems, have been extensively used in studies to characterize the appearance of pain in the human face. However, its application cannot be done in the practice in real-time, since it is burdensome and requires multiple observations of behavioral samples to identify the actions of separate muscle groups.1
Recently, other tools have been developed to assess the facial expression of pain. For instance, the Critical-Care Pain Observation Tool assesses factors such as relaxation, tension, and grimacing. However, this is usually applied by nurses, thus increasing the workload for health personnel.3
Several investigations have made attempts to develop a series of automatic recognition of facial expressions of pain, including classifier models based on support vector machines and fusion network architectures. However, its application in intensive care settings is difficult, since it is necessary to obtain standardized and complete facial images in hospitalized patients who may have an endotracheal tube, a nasoesophageal tube, or an oxygen mask. In addition, facial muscle movements associated with pain may be difficult to notice due to sedation and tissue edema.3
To address this need, Wu and colleagues published in March 2022 the development of a classifier based on deep learning that determines pain from videoclips of facial expressions in critically ill patients.3
The study was conducted with 63 patients admitted to medical and surgical intensive care units, 81 percent of them in critical condition. Video recording and labeling were performed by three experienced and trained nurses, while labeling was validated by two senior nurses.3
To mitigate information bias, the protocol required to observe the patient for 10 seconds, to record a video for 20 seconds, and then the pain score was labeled at the end of the video. The recordings were made particularly before and after suction, dressing changes, and invasive procedures, since the aim was to obtain the videos with different degrees of pain.3
To address the problem of facial interpretation in intensive care patients, a facial landmark tracker was used to locate the facial area. This tracker was especially focused on the area near the eyebrow, which is less likely to be masked by medical devices. In addition, a convolutional neural network was responsible for locating the facial area if the tracker failed to locate the face.3
Afterwards, three tagged videos were recorded for each observation, giving a final total of 746 videos eligible for analysis. The researchers then established two types of classifiers, one for image and one for video, using convolutional neural network models. These included Resnet34, VGG16, InceptionV1, and bidirectional short-term memory networks.3
In the participating hospitals, the Critical-Care Pain Observation Tool is used as a standard of care, so the pain scores based on facial expression are in accordance with its criteria. Pain intensities were defined as a zero-score, equivalent to a relaxed expression; a one-score, equivalent to a tense expression, and a two-score or grimacing, which reflects clinical alert signaling that requires immediate attention.3
Researchers used two types of classifiers: the first was polychrome, which attempted to classify patients into the three pain categories, being these “zero”, “one” and “two”. The second was dichotomous, which attempted to classify two pain categories by comparing the zero-score category with a set that incorporated the categories one and two, and further comparing the categories zero and two.3
The results showed that the accuracy of the dichotomous classifier in differentiating tense/grimacing from relaxed facial expression was 80 percent, and the accuracy in detecting grimaces was almost 90 percent. Furthermore, the performance of the video classifiers was better than that of the image classifiers. This detail reveals how important the temporal relationship between image frames is for classifying pain by facial expressions.3
To conclude, the study performed by Wu and colleagues demonstrates a practical application of deep learning-based automated pain assessment in the intensive care unit. The application of these findings might improve both quality of care and routine workload in emergency settings. This may be especially relevant in the post-COVID era, where contactless monitoring in intensive care units is becoming essential.3
Capsule
It is estimated that more than 50 percent of patients in the intensive care unit experience moderate to severe pain at rest, while 80 percent of those in critical condition experience pain during procedures. Increased pain has been linked to anxiety, delirium, and poor outcomes in both the short and long term. Therefore, severe pain may reflect not only inadequate pain control but also patient deterioration caused by the disease. Several studies have shown that regular pain assessment is associated with better outcomes, including time spent on the ventilator.3
End of capsule
Welcome back. In the previous segment, we discussed behavioral pain indicators and how they can be used to build automated pain assessment tools. We now turn to physiological indicators of pain. This category includes autonomic responses such as electrodermal activity, oxygen saturation, heart rate variability, evoked potentials, as well as activation of certain brain regions. Assessment of these automated responses might help identify and quantify pain.1
A growing number of studies have investigated machine learning models based on physiological signals for objective estimation of pain intensity, but have not yielded estimates of acceptable precision in clinical settings. One of the main challenges in achieving automatic assessment is the significant variability between individual perceptions and physiological responses to pain.2
An interesting approach to this problem is the one published in July 2021 by Fatemeh Pouromran and colleagues. Seeking to investigate whether physiological signs can replace or complement a patient’s self-report, the team evaluated the performance of different machine learning models, several time-series features, and a series of physiological sensors. With all of these parameters, the authors succeeded in building a computationally inexpensive model that can estimate objective pain.2
An external database was used for this study, which includes 87 healthy people in the age range of 18 to 65 years. These participants were subjected to heat pain in the right arm with four subject-dependent temperature intensities. Then, the research team recorded electrodermal activity, electrocardiogram, and electromyogram signals.2
It is interesting to understand why these physiological signals are potential candidates for the development of machine learning pain measurement. The pain process often begins with the activation of the sensory neural pathway by noxious stimuli that can be either internal or external. Then, the consequent activation of the autonomic nervous system triggers physiological changes.2
When a painful stimulus occurs, electrodermal activity measures the changes in the electrical properties of the skin that occur when the sympathetic nervous system innervates the sweat glands to secrete sweat. Similarly, the electrocardiogram records the increase in heart rate and the changes in its variability and low-frequency. Likewise, electromyogram records the electrical activity of the muscles, particularly in the trapezius muscle, which is considered an indicator of sympathetic and parasympathetic nervous system activation during physiological arousal.2
With all these characteristics in mind, the researchers built different machine learning models based on the features extracted from the electrodermal activity, electrocardiogram, and electromyogram signals, in order to evaluate their performance for the continuous estimation of pain intensity. These models included linear regression, support vector regression, neural networks, and extreme gradient boosting.2
The models were developed based on two pain estimation scenarios: First, a subject-independent model, which is a generic model for all participants, where the pain intensity of a new subject is calculated based on the patterns discovered in samples from all other subjects. The second is a subject-dependent model, dedicated to each person individually. The model is trained and tested on different data samples from the same person, then a repeated 10-fold stratified cross-validation method is applied. The performance metrics used were mean absolute error and root-mean-square error.2
The reported findings suggest that electrodermal activity is the best signal for estimating pain intensity for both subject-independent and subject-dependent scenarios. Using only 3 simple time-series features, it obtained a mean absolute error of 0.93. On the other hand, among all tested models, support vector regression had the best predictive performance across different sensors and features.2
Finally, the team tested a specific hybrid cluster model that combines the advantages of subject-independent and subject-dependent models. To do this, a cluster-based support vector regression model was built to address the variability in physiological responses to pain.2
This model can provide personalization without the need to build an individual model for each subject. Each signal was tested separately to find the ones that benefited from the hybrid approach. Electrocardiogram and electromyogram signals showed inter-subject variability in responses to pain, which affects the automated models' ability to generalize across people. Meanwhile, the electrodermal activity signal was once again the superior candidate.2
The study concludes that it is possible to estimate target pain using only the electrodermal activity signal, which does not need a complex setup or a demanding algorithm. The research team proposed that these findings could lead to the development of a wearable pain measurement device for online monitoring that could increase the quality of care that patients receive.2
Looking to the future of this research, the authors point to some of the challenges that remain underexplored in the field of pain assessment. First, they recognize that the physiological signals involved in responses to pain can be affected by emotion, anxiety, sleep, and other experiences.2
It is worth mentioning that, although physiological measures are used routinely in both clinical and research settings, several factors limit their use: First, as Pouromran and colleagues mention,2 it is possible that these measures covary with other affective states that frequently are correlated with pain, such as fear.1
Likewise, some measures serve as indices of processes that are influenced by pain rather than the pain itself. Also, it is not uncommon that physiological measures result hyperreactive, or on the contrary, have an insufficient response depending on each patient; so, the discrimination between pain states is poor. In addition, most patients require a physiologic evaluation that is moderately invasive or involves special equipment.1
On that account, it is critical to have models that can adequately distinguish the reason for these changes in physiological signals. Moreover, physiological responses to pain may be different in healthy subjects and patients with specific diseases, so having models calibrated for specific groups should be a research priority.2
Thanks for joining us on this episode of Health Connect. Don't miss our next episode, where we will talk about more artificial intelligence developments in other medical fields. Discover more medical news and content on Viatris Connect.
References:
- 1. Prkachin KM, Hammal Z. Computer mediated automatic detection of pain-related behavior: prospect, progress, perils. Front Pain Res (Lausanne). [Internet] 2021. (Accessed on Jun 8, 2022);1;2:788606. Available at: https://www.frontiersin.org/articles/10.3389/fpain.2021.788606/full
- 2. Pouromran F, Radhakrishnan S, Kamarthi S. Exploration of physiological sensors, features, and machine learning models for pain intensity estimation. PLoS One. [Internet] 2021. (Accessed on Jun 8, 2022);16(7):e0254108. Available at: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0254108
- 3. Wu CL, Liu SF, Yu TL, Shih SJ, Chang CH, Yang Mao SF, Li YS, Chen HJ, Chen CC, Chao WC. Deep Learning-Based Pain Classifier Based on the Facial Expression in Critically Ill Patients. Front Med (Lausanne). [Internet] 2022. (Accessed on Jun 8, 2022);9:851690. Available at: https://www.frontiersin.org/articles/10.3389/fmed.2022.851690/full
Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.
NON-2022-13466
NON-2023-2512