3.1.1. Facial Gestures (Grimace, Mouth Opened, and Raising of Eyebrows)

Using a program with AI for face recognition, the smartphone camera captures facial movements. Facial expressions are a key in expressing emotions and recognizing pain. Hence, the majority of research has concentrated on this approach. The introduction of the NBC-McMaster database, the first public database for pain recognition, further strengthened this trend by offering detailed annotations along with facial images. Alternatively, other stress-inducing factors can induce a condition like pain, resulting in simulated facial expressions of pain [16].

Facial expressions are a helpful tool for determining pain levels as pain tends to be connected with natural facial movements. Simple pictures or videos of faces were frequently utilized in this context. The FACS is a technique used to describe and study observable facial expressions. It consists of 44 basic units known as action units (AUs) that correspond to the muscle activation and evaluate movements. FACS offers a manual with a collection of pictures and drawings to assist in scoring these AUs, offering guidance. Facial animation software is utilized in psychology, sociology, and communication studies, as well as for facial palsy assessment.

Convolutional neural networks (CNN) are used for image classification and object recognition tasks. CNNs are a form of feed-forward artificial neural network (ANN), in which there are no loops created by connections between nodes. They perform the identical role to the retina by transforming input information into output. It comprises multiple layers, starting with the convolutional layer, followed by a nonlinear layer and several pooling layers before reaching the fully connected layer (flattening). Classification is carried out in the final layer by using the features obtained from earlier layers and different filters. The CNN structure may differ, but the VGG 16 architecture serves as a model for constructing CNNs. It is made up of 16 convolutional layers with a size of 3 × 3 and several filters. VGGFace is frequently utilized in pain research. This model has over 3 million face images of over 9000 individuals and is a version of the VGG16 and VGG19 models, utilizing a vast dataset of facial images (VGGFace2). The VGGFace model is suitable for tasks such as face recognition, face verification, and emotion recognition. In this regard, Zamzmi et al. used a novel lightweight convolutional neural network as well as other popular CNN architectures to assess infant pain. They suggest that automated infant pain detection using CNNs is a suitable and more efficient alternative to traditional assessment [17]. Hosseini et al. also used a deep CNN (DCNN) model through facial expressions to predict pain intensity, which yielded promising results with high accuracy [18]. Bargshady et al. also used a Visual Geometric Group Face Convolutional Neural Network (VGGFace CNN) to classify pain intensity from face [19].

Various kinds of ANNs can also be applied to tasks involving image processing. For instance, deep residual networks (ResNets) enable enhanced performance. Additionally, generative adversarial networks (GANs) are a form of ANNs comprising two neural networks: one creating images and the other determining authenticity by comparing them to real images. The autoencoder is another type of neural network architecture utilized for compressing and reconstructing images. U-Nets are another type of CNN model that is employed for tasks involving the segmentation of images. For example, Monwar and Rezaei used efficient video analysis techniques and ANNs to classify painful and painless faces and pain recognition [20]. Shuang et al. also used ResNet50 to build an automatic pain expression recognition model in elderly patients with hip fractures, which had high accuracy [21]. Wang and colleagues also used a generative deep learning method to predict pain in patients in the ICU, which had an accuracy of 97.2% [22].

RNNs are neural networks created for the purpose of handling sequential data like time series information or natural language text, enabling the transfer of information between different stages of the sequence. LSTM, a form of RNN, excels at recognizing and retaining long-term relationships in sequential data. LSTMs have been applied to different tasks like processing natural language, recognizing speech, and predicting time series data. BiLSTM is a type of network that includes two LSTMs, with one handling input in the forward direction and the other handling input in the backward direction. They have the ability to examine data including facial expressions, body language, speech, and heart rate (HR). CNNs excel in extracting features from images, whereas LSTM networks excel in handling sequential data. Thus, a hybrid neural network is suggested for pain research, which integrates both a CNN and an LSTM network. Cascella et al. recently employed CNN and LSTM to categorize facial emotions from diverse openly accessible datasets [9, 16]. The results of the study by Rojas et al. showed that the BiLSTM model had the highest accuracy in assessing pain in patients who were unable to speak [23].

CNN and ANN are both types of neural networks used in ML, but they have different architectures and are suited for different types of tasks. When it comes to facial expression recognition, CNNs are generally considered to be more effective due to their ability to capture spatial hierarchies and patterns in images. The differences between CNN and ANN in the context of facial expression recognition are outlined as follows (Table 2).

3.1.2. Which Technique Is Better for Facial Expression Recognition?

In summary, CNNs are the preferred choice for facial expression recognition due to their ability to handle image data effectively and their superior performance in capturing the intricate details necessary for accurate expression classification.

3.1.3. Example of “Pain” vs. “No-Pain” States in Camera-Based Approaches

In the context of pain management in cancer patients, camera-based approaches often involve the use of facial recognition and analysis to distinguish between “pain” and “no-pain” states. These systems typically rely on the detection of facial expressions associated with pain, which can be challenging due to the subjective nature of pain and the potential for variations in individual pain responses [24].

3.1.4. Guidelines for Differentiating Pain States

3.1.5. Incorporation Into Pain Management

Camera-based approaches offer a promising avenue for objective pain assessment in cancer patients. By distinguishing between “pain” and “no-pain” states through facial expression analysis, these systems can enhance the accuracy and timeliness of pain management interventions. However, it is important to note that these approaches should be used in conjunction with other pain assessment methods and patient self-reports to ensure comprehensive and effective pain management [26].

3.2.1. Physiological Measures

Electrodermal activity (EDA), surface electromyography (sEMG), respiration (RESP), oxygen saturation (SpO2), blood pressure (BP), electrocardiogram (ECG), movement (MOVE), skin temperature (SKT), pupillary response (PUPIL), and photoplethysmography (PPG) are all physiological measurements that can be used in research and monitoring.

EDA is a widely utilized physiological signal in psychiatry and psychology research. EDA, also known as galvanic skin response or skin conductance response, measures changes in the electrical resistance or conductivity of the skin. This signal reflects changes in sweat gland activity, which is modulated by the sympathetic nervous system. Increased sweating in response to stimuli leads to a decrease in skin resistance and an increase in skin conductance. EDA is commonly recorded from the palms and soles of the feet due to the higher density of glands in these areas [27].

Another commonly used method to evaluate pain is electrocardiography (ECG). The ECG signal shows the heart’s contracting activity. The autonomic nervous system is responsible for controlling the internal functions of the body, such as the cardiovascular system, without conscious effort. The most crucial pain-related parameters of ECG signals are HR and heart rate variability (HRV). Pain is linked to higher HR and lower HRV [27].

Electromyography (EMG) serves as an alternative method for pain assessment. This method evaluates muscle electrical activity on the skin’s surface. No specific muscle serves as a pain indicator, but heightened trapezius activity is associated with stress in affective computing, whereas increased zygomaticus and corrugator muscle activity is linked to reacting to negative images. The trapezius muscle, zygomaticus major, and corrugator supercilii muscles were utilized for EMG measurement in the examined articles [27].

PPG is another reliable technique for detecting alterations in blood volume and is frequently utilized in transmission (like a clip on an earlobe or finger) or reflective mode (as seen in a wristwatch). The finger or earlobe is where most often seen. The PPG signal allows for the determination of metrics like HR, SpO₂, and respiratory rate.

SKT is an additional metric for detecting pain. The skin’s resistance decreases when sweat is produced. Because there is a greater concentration of glands in the palms and soles, this characteristic can be assessed in those regions.

Other signs like RESP rate, SpO₂ levels, BP readings, MOVE patterns, and changes in pupil size (PPIL) are also utilized to assess pain. Respiratory alterations frequently happen due to pain and can be monitored with the use of a stretchy band placed around the chest. Changes in breathing, oxygen levels, and BP are linked to pain. Painful stimulation also led to a decrease in the size of pupils, also known as dilation.

Most contact-based sensors have drawbacks: movement can cause errors and the signals can be influenced by changes in the sensor’s contact with the skin, such as adhesive electrodes becoming loose and needing reattachment [27, 28].

3.2.2. Neurological Measures

The human brain plays a crucial role in responding to stimuli, as neural signals are interconnected with intricate functions like sensory and motor coordination. In research, two different neurosensors were employed for pain evaluation: electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS).

EEG was the most frequently used method for assessing pain. EEG is used to monitor the electrical signals in the brain, which are analyzed to evaluate levels of pain. EEG is the most commonly used method to evaluate cognitive functions like attention and alertness. EEG data are segmented into five categories: delta, theta, alpha, beta, and gamma. Evaluations make use of the variations in the intensity of each of these bands.

Another method for evaluating pain using neurosensors involves assessing the activity of various brain areas through fNIRS, which detects levels of oxygenation (HbO) and deoxygenated hemoglobin (HbR) in the brain cortex [27].

5.1.1. Predictive Analytics in Cancer

Large amounts of data and predictive analytics can greatly enhance risk assessment, particularly in data-heavy areas such as oncology. A field abundant in data, such as oncology, appears to be a suitable fit for predictive analytics [32]. In the past few years, there has been a significant increase in the field of AI within oncology. AI solutions have been created to address various challenges related to cancer. Healthcare facilities, hospitals, and tech firms are creating AI tools to help with medical decision-making, enhance cancer treatment accessibility, and boost clinical productivity in order to provide secure, top-quality cancer care. AI in the field of oncology has shown high levels of accuracy in tasks such as analyzing images, making predictions, and providing precise treatment options for cancer patients [33]. Predictive analysis tools use historical patient data to automatically project future health results for individuals or groups. As the volume of data in oncology, including EHR, radiology, genomic, and other types, has grown, there have been multiple potentially applicable use cases [32].

5.1.2. Predictive Analytics in Oncology: Challenges

Despite the necessity for improved predictions of life expectancy, acute care utilization, adverse effects, and genomic and molecular risk, there are few instances of predictive analytics being utilized in the field of oncology. We believe that existing predictive analytic interventions have the potential to fill significant gaps in risk stratification strategies in the field of oncology. Nevertheless, healthcare professionals, technology experts, and decision-makers need to tackle the obstacles related to research, technology, and regulations that hinder the use of analytics in the field of oncology [32].

Risk assessment in the field of cancer treatment is hampered by the absence of necessary prognostic details, the requirement for laborious manual data entry, limited access to thorough data, and in certain instances, an excessive dependence on clinical judgment. Lack of accurate identification of high-risk patients with cancer can result in either overly aggressive end-of-life treatment or unnecessary use of acute care. Additionally, the evaluation of prognostic factors such as performance status can be influenced by variations and prejudices among clinicians [32].

5.1.3. Predictive Analytics and Molecular Markers

Recent research is exploring how tumor genomic and molecular biomarkers can predict the success of treatments and patient outcomes. Currently, there are minimal studies that have tried to combine genomic and clinical biomarkers in predictive models. With the introduction of molecular-based treatments and immunotherapy, cancers must be categorized more specifically according to these molecular indicators. Newer systems like the updated Katagiri have integrated molecular features of cancer in the prognostic model, including the hormone status of breast cancer and molecular targets of non–small-cell lung carcinoma. Considering genetic or molecular expression in predictive analytics may represent the next frontier in predicting clinical outcomes [34]. Combining information from clinical, genetic, and molecular predictive factors could lead to a more accurate and detailed risk assessment in oncology, paving the way for precise precision medicine [32].

5.1.4. Radiomics

The growing area of radiomics is an example of how predictive analytics models are starting to be utilized in the field of oncology. Radiomics is the study of tumor characteristics through quantitative data analysis of scans. These traits can help in identifying, describing, and keeping track of solid tumors. Computer-aided detection can be used to detect cancerous lung nodules on CT scans and prostate lesions on MRI scans, and it might also be useful for automating tumor staging [32].

5.1.5. Pathology

Pathology is a crucial area in the practice of oncology that is ready to see advantages from predictive analytics. There are many models developed to predict outcomes and anticipate therapy responses based on tumor pathologic characteristics for certain diseases. The 21-gene recurrence score and the 70-gene recurrence score are commonly used in clinical practice for determining the need for chemotherapy in patients with early-stage breast cancer [32].

5.1.6. Predictive Computer-Based Models

Experimental and computational models of cancer have offered a valuable option to studying patients directly, helping to address concerns related to experimental design, ethics, finances, and animal welfare. The results can be related to in vivo studies to enhance the models and aid in designing predictive-focused experiments (Figure 3) [35].

5.1.7. Utilize Predictive Analytics in Pain Management

Dealing with pain is seen as a worldwide issue that needs a combined effort due to its complex nature, involving various aspects such as biology, culture, development, psychology, empathy, and situational factors. By examining extensive data on pain treatment from vast databases like EHRs and information systems, it is feasible to offer real-time guidance on care management and identify patient characteristics through their clinical data repositories, leading to more precise decisions in this area in Figure 4 [36].

5.2.1. Telehealth Applied to Cancer Patients

Improvements in cancer diagnosis and treatment have led to a rise in cancer survival rates. However, cancer survivors often face issues related to their disease or treatment, resulting in a lower quality of life. Different telehealth interventions have been widely used in medical care, proving to be cost-effective, highly satisfying for patients, and well-accepted by health professionals. Telehealth interventions are a successful alternative for enhancing the quality of life in cancer survivors [38].

Successful provision of cancer care through telehealth necessitates a well-thought-out system that considers various resources such as patients, healthcare providers, and clinic/cancer center resources before, during, and after the treatment. Utilizing telehealth for cancer care involves more than just substituting in-person appointments with audio or video consultations; it necessitates thorough program design and proactive education for patients, providers, and healthcare facilities. Even though telehealth seems to enhance fairness in accessing healthcare, persistent challenges related to poverty, underinsurance, limited technology and broadband access, and health literacy continue to hinder equitable care access in low-income areas [37].

5.2.2. Telehealth in Cancer Pain Management

In patients with cancer pain, using telemedicine strategies can improve clinical care and make better use of resources. This method has the potential to play a crucial role in enhancing patient satisfaction and enhancing health results [40].

Nevertheless, addressing pain associated with cancer usually necessitates a multifaceted and varied strategy. Telemedicine has the potential to provide a range of options for reassessing paths of treatment, such as managing cancer pain. However, even with the increasing utilization of telehealth approaches, there is still a lack of scientific proof to create care plans [40]. Currently, there is insufficient evidence regarding the potential uses of telemedicine in helping cancer patients manage pain [9].

6.1.1. The Role of Predictive Analytics in Opioid Crisis Prevention

Due to the opioid crisis, it is crucial to accurately assess overdose risk and enhance knowledge on providing effective treatments and interventions for individuals with opioid use disorder to minimize overdose risk. Rises in opioid overdoses from pharmaceuticals highlight the importance of supporting prescribers with these tools for prevention planning. A thorough public health response based on evidence is necessary to lessen the negative effects of opioid use. Expanding current interventions, treatments, and resources necessitates a focused strategy to determine the best combination and level of interventions. Although there is proof that some interventions can reduce the risk of overdose, further research is required to determine the most effective way to implement these interventions. Recently, there has been a growing fascination with employing ML techniques to create prediction models. These techniques are frequently touted as being more effective than conventional regression because of their reliance on data and capacity to recognize patterns in data (Table 5) [44].

6.2.1. NSAIDs

This set of medications produces pain-relieving effects on both the central and peripheral nervous systems [42]. Additionally, they can be combined with opioids as a supplementary method for pain management to ensure effective pain relief without interfering with cognitive function or inducing sedation. Nonetheless, the use of NSAIDs comes with various risks, including renal impairment, GI bleeding or perforation, thrombocytopenia, bleeding disorders, and cardiac toxicities [47]. Combining NSAIDs with acetaminophen is beneficial, as the effectiveness of the mixture surpasses that of each individual drug [42]. Cancer patients could face an increased likelihood of experiencing negative effects from NSAIDs due to their cancer treatments or the disease itself, which may already make them more susceptible to such problems [47].

6.2.2. Anticonvulsants

Anticonvulsants like gabapentin and pregabalin might help individuals with neuropathic pain and have the potential to lower the need for opioids, resulting in fewer negative effects [47].

6.2.3. Antidepressants

Medications like duloxetine and venlafaxine, which are serotonin–norepinephrine reuptake inhibitors, as well as tricyclic antidepressants like amitriptyline, desipramine, imipramine, and nortriptyline, are options for managing neuropathic pain [47].