Estimation of Pain Characteristics through Voices Using Artificial Intelligence Technology

Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. Today, we will talk about pain measurement through voice-based artificial intelligence models. 

Did you know that an accurate pain estimation is critical for effective pain management? Improvements in its precision can reduce the risk of incorrect treatment.¹

Pain is a vital feedback mechanism in the human body that functions to keep the body in homeostasis. It signals the body in response to damage and aids in the prevention of further illness.² 

Although pain is a universal sensation, pain and its chronification are still not completely understood and remain unresolved medical issues.²

Chronic pain is the most prevalent chronic disease across the globe. It negatively impacts the quality of life and function, affecting individuals, their families, communities, and health systems. However, the true burden lies in the proper assessment of pain and outcomes with numerous modalities of treatments utilized.²

Currently, the gold standard of pain measurement is self-reporting scales. However, these methods of evaluation are subject to high variability in individual perceptions of pain.2 In addition, since these strategies depend on patients’ accounts, they are exclusively applicable to individuals with no verbal or cognitive impairments.¹ 

Thus, all these established techniques are inapplicable to newborns, delirious patients, patients under sedation and ventilation, or individuals with dementia or developmental and intellectual disorders. Such patients are entirely dependent on others’ awareness of nonverbal pain cues.¹

In light of this, observational pain scales are advised for use among adults in critical conditions when self-reported pain measurements are not feasible. However, the reliability and validity of these measurements are still limited, because even qualified raters cannot ensure an unbiased judgment.¹

Despite the thorough understanding of the pathophysiological processes underlying the physical pain response and the technological advances to date, pain is often poorly managed. Misdiagnosis of pain levels through associated subjective biases can increase unnecessary costs and risks. In addition, poor pain relief can result in emotional distress and is linked to several consequences.¹

Although breakthroughs in computing have previously been observed regarding medical applications, it can be challenging to distinguish between various types of pain. Chronic pain is more difficult to diagnose and necessitates the recognition of more subtle pain coping mechanisms, but the accompanying facial, behavioral, and verbal pain indications make it quite simple to recognize acute pain.¹

Automatic pain recognition has transitioned from being a theory to becoming a highly important area of study in recent times. As a result, a few studies have concentrated on utilizing artificial intelligence to identify or classify pain levels using audio inputs.¹

Artificial intelligence (or AI) describes a system’s ability to mimic human behavior and exhibit intelligence, which is now viewed as a branch of engineering.¹

Although the evolution of AI can be traced back to 1956, the concept really began to flourish in the last decade. This is due to a huge boost in AI performance and speed with the help of low-cost high-performance graphics processing units, cheaper storage options, faster cloud computing, parallel computing advancements coupled with a surge in data collection.²

Thus, artificial intelligence combines computer science, engineering, and related disciplines to build machines capable of behavior that would be said to require intelligence were it to be observed in humans.² 

Such behaviors may be described as the ability to visually perceive images, recognize speech, translate language, and learn from and adapt to new information.²

There are two main categories of AI applications in the field of medicine: virtual and physical. The virtual subfield involves machine learning, natural language processing, and deep learning. On the other hand, the physical subfield involves medical equipment and care bots, which assist in delivering medical care.¹

The operation of artificial neurons is governed by three fundamental principles: multiplication, summation, and activation. Initially, each input value is multiplied by a distinctive weight. Subsequently, a summation function aggregates all the weighted inputs. Finally, at the output layer, a transfer function transmits the sum of the previous inputs and bias.¹

According to some studies, verbal or acoustic cues can be used to identify and even categorize pain according to its intensity level. The mean and range of voice fundamental frequency, loudness, and degree of periodicity are some of the acoustic characteristics of voice that have been shown to indicate pain levels.¹

Pain-related vocalizations, such as screaming, groaning, moaning, and crying, are often accompanied by changes in these phonetic features and can also be suggestive of pain.¹

Researchers have begun to create algorithms to automate pain level assessment using speech thanks to developments in signal processing and machine learning methods.¹

To extract speech-specific features from the recorded voice, specific analysis techniques are required. These techniques can be used for various tasks such as speech recognition, speaker identification, and emotion recognition.¹

To be able to use AI models in pain detection from voice, multiple artificial neurons are merged to create an artificial neural network. These artificial neurons imitate the behavior and structure of biological neurons. Furthermore, to enhance the efficiency and precision of the artificial neural network, the neurons are organized into layers for ease of manipulation and precise mathematical representation.¹

The topology of an artificial neural network refers to how different artificial neurons are coupled with one another. Different artificial neural network topographies are appropriate for addressing different issues.¹

Once the topology has been selected and it has been tuned and learned the right behavior, an artificial neural network can solve a problem.¹

The method of “embedding stacked bottleneck features”, a specific type of deep neural network architecture, provides a major advantage in audio processing. This method is computationally more efficient and significantly enhances the performance of AI models when confronted with complex data containing numerous variables.¹ 

In 2017, Tsai and colleagues conducted one of the first and most influential studies in the field of AI used to detect pain with voice in real adult patients. They employed a stacked bottleneck long short-term memory model, intending to classify the patients’ pain levels. Patients who initially presented with symptoms of pain in chest, abdomen, limbs, and headaches had their Numerical Rating Scale pain scores recorded for the follow-up period. During these sessions, nurses conversed with the patients regarding their pain characteristics and location. Prosodic and spectral elements were extracted from the vocal data to create two different subgroups.¹

The authors created a compressed data representation with reduced dimensions and utilized stacked bottleneck acoustic attributes with long short-term memory to generate deep bottleneck features. Then, they fine-tuned the stacked bottleneck layer by encoding the patients’ sentences through different network layers. Lastly, they used the support vector machine to classify the pain levels. In this study, they showed that this novel approach could obtain a good accuracy in binary and tertiary pain classification tasks.¹

Cápsula
The application of artificial intelligence has great potential to improve healthcare and its use will continue to grow in this field for tasks such as screening and evaluation, and, while it should not replace clinical judgement, it could help clinicians make better decisions.³

However, there is a continuous debate regarding whether AI “fits within existing legal categories or whether a new category with its special features and implications should be developed.”³ 

To fully achieve its potential in healthcare, several measures need to be put in place to regulate AI and keep it morally accountable. These measures include data privacy, safety, transparency, unbiased algorithmic fairness and inclusive programming, but perhaps more importantly, frequent audits.³

Fin de la cápsula.

In 2021, the main objective of the study conducted by Mohan and colleagues was to design an audio-based intelligent system to recognize distress signals during myocardial infarction in an interior environment with a low-power embedded graphic processing unit edge device coupled with deep learning.¹

This device was small and portable. It included a graphic processing unit for conducting accelerated computing operations at a network’s edge rather than in a centralized data center or cloud environment. The authors instructed 60 people to scream and communicate as if they were having a heart attack.¹

The study included three stages: In stage 1, audio data samples were gathered and prepared. A Convolutional Neural Network model was trained during stage 2, and stage 3 involved the deployment of the model trained. Finally, the researchers also tested the Convolutional Neural Network model on a separate, private audio dataset to validate their approach. Overall, the study results suggest that the proposed supervised audio classification scheme can effectively detect distress sounds in an indoor smart space, even in challenging environments with background noise.¹

On a different note, research has shown that ethnicity, race, gender, and age may all impact how someone experiences pain. For example, women tend to encounter more intense pain that occurs more frequently and lasts longer than men. While evidence suggests a decline in pain perception as individuals age, older adults may experience heightened sensitivity to pressure-induced pain.¹

Li and collaborators proposed a method for automatic pain level recognition that involves the learning of latent acoustic representations embedded with attributes of gender and age. They achieved this by utilizing a maximum mean discrepancy variational autoencode. For the preprocessing task, the authors utilized the acoustic low-level descriptors set called the extended Geneva Minimalistic Acoustic Parameter Set.¹

The features were subjected to z-normalization for every speaker, known as speaker normalization. Subsequently, context window expansion was implemented. The acoustic low-level descriptors were also transformed into hidden features using a conditional variational autoencoder.¹

Typically, personal attribute dependency in pain recognition is addressed by training multiple independent models, such as gender-specific or age-specific models. However, the authors’ proposed approach of maximum mean discrepancy variational autoencode directly embeds this information in the encoded acoustic space. The authors demonstrate that, compared to females, male patients with severe pain express their pain with limited maximum jitter. However, female voices are more variable in pitch and have less harmonic content than male voices. In addition, elderly patients’ voices, in the context of pain, are less variable in pitch and amplitude than non-elderly voices.¹

Deep learning algorithms have a significant advantage in representing complex problems and subjective measurements, such as pain. However, the majority of the research in this field has focused on exploiting visual clues, including facial expressions, to identify pain.¹

Even though vocalizations are a well-recognized sign of pain and are frequently used in clinical contexts to measure pain intensity, the application of artificial intelligence with voice or audio to identify pain has received much less attention. The difficulties in interpreting voice or audio data may be one factor affecting the limited number of studies in this area.¹

The pitch, length, and other acoustic characteristics of pain vocalizations can vary greatly, making it challenging to create robust algorithms for pain detection based on these signals. Additionally, it can be challenging to separate pain vocalizations from other voices due to background noise and other confusing elements. ¹

In studies that explore the detection of pain using artificial intelligence, the accuracy of pain measurement could be improved by creating a combined multimodal framework that incorporates facial expressions, body language, biological signals, audio, and vocalization data. ¹

Despite the challenges, there is growing interest in the utilization of speech or audio with AI to detect pain, as it could have a wide range of clinical and research applications. However, more investigation is necessary to understand the factors that affect the correlation between vocalization and pain.¹

Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode.

Discover more medical news on Viatris Connect.

References:

1. 1. Borna S, Haider CR, Maita KC, Torres RA, Avila FR, Garcia JP, et al. A Review of Voice-Based Pain Detection in Adults Using Artificial Intelligence. Bioengineering. 2023;10:500. Available at: https://www.mdpi.com/2306-5354/10/4/500
2. 2. Nagireddi JN, Vyas AK, Sanapati MR, Soin A, Manchikanti L. The Analysis of Pain Research through the Lens of Artificial Intelligence and Machine Learning. Pain Physician. 2022;25(2):E211-E243. Available at: https://www.painphysicianjournal.com/linkout?issn=&vol=25&page=E211
3. 3. Naik N, Hameed BMZ, Shetty DK, Swain D, Shah M, Paul R, et al. Legal and Ethical Consideration in Artificial Intelligence in Healthcare: Who Takes Responsibility? Front Surg. 2023;9:862322. Available at: https://www.frontiersin.org/articles/10.3389/fsurg.2022.862322/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-2023-9800
NON-2023-2512