According to the International Association for the Study of Pain, pain is an unpleasant sensory and emotional experience associated with, or resembling that is associated with, actual or potential tissue damage [1]. One common distinction is between acute and chronic pain. Acute pain is typically caused by the stimulation of peripheral nerve endings because of inflammation or trauma or interference with nerve pathways (neuropathic) because of the nerve being severed (for example, during surgery). Tissue healing generally results in the cessation of acute pain. Chronic pain arises in acute pain situations, generally when the acute pain is very intense or lasts for an extended period. It is important to remember that in most situations, acute pain serves as a critical protective mechanism in preventing further tissue injury. By reducing the risk of continued trauma, the tissue can heal more rapidly and the pain will subside. The main challenge in dealing with intensive acute pain is to prevent the overuse of strong opioids, morphine, codeine, and cocaine, leading to addiction through the euphoria created using these medications.

According to the National Health Interview Survey of 2019, a total of 50.2 million or approximately 20.5% of American adults experience chronic pain, with the most common examples being back pain and hip, knee, or foot pain [2]. Chronic low back pain affects a significant segment of the population. It is a heterogeneous disease that includes several causes of pain syndromes, latent molecular pathologies, genetic and psychological factors, and a history of injury. The Institute of Medicine has estimated that chronic pain affects approximately 100 million adults in the United States, with an estimated annual cost of up to US $635 billion [3].

Pain is perceived centrally and is strongly influenced by physical, physiological, and social or cultural factors. Another distinction often made is between pain caused by tissue damage (sometimes called inflammatory or nociceptive) and pain caused by nerve damage (neuropathic), where nerve signals may not be driven by local tissue damage.

The approach we propose is to devise a BN that will consider what evidence is available and report how the available evidence may be interpreted as a distribution of the likelihood that the participant is experiencing different levels of pain. This is similar to the proposal by Hill et al [59]. As a starting illustration, we consider only three forms of evidence: (1) a participant self-reports pain on a 0 to 5 scale (Figure 3), (2) a measured value of serum levels of COX-2 on a similar 0 to 5 scale, and (3) a measured value of inducible nitric oxide synthase (iNOS) on a 0 to 3 scale.

The process of developing a BN involves 2 basic steps:

where P(A) and P(B) are the prior probabilities, and P(A|B) is the conditional probability of A given the level of B. The usual approach assumes that all levels of a node’s priors are the same (the principle of equal ignorance). However, in some applications, we may have the knowledge that the priors are not the same, and we can use this knowledge. If B is a set of parents rather than a singleton, a chain rule applies. Suppose it is the parent that is unknown (here, B is the child and A is the parent; in Figure 4, experienced pain is the parent and reported pain is the child). In this case, we compute the posterior probabilities for each level (i) of the parent node using the following (equation 2):

Each node in a specified graph requires a prior probability estimate. As mentioned earlier, one may always assume (by the principle of equal ignorance) that each level of a node has equal prior probability. If we have the knowledge or even reasonable consensus from experts that some probabilities other than all-equal are better, say, for a particular application or population of patients, we may set them accordingly. Such assumptions can always be tested using the methods described next.

The difficult challenge in devising a BN for an application is specifying the conditional probabilities for the parent-child links. One rarely has sufficient knowledge at this fine level of detail. However, what we may be able to accomplish with an adequate pool of domain experts is a specification of “reasonableness” tests for the final probabilities. That is, we ask the experts to say how the probabilities for the different levels of experienced pain (which we can never know with any certainty) “should” come out for some set of test cases. This set of test cases should span a “representative” range of possibilities of evidence combinations. With such information, it should be possible to set up an optimization procedure that can set the required conditional probability parameters to closely approximate the desired output behaviors of the system. This approach is illustrated in Figure 5.

Our earlier study explored the potential utility of serum COX-2 and iNOS as objective measures of pain in 102 American patients [60]. Sandwich enzyme-linked immunosorbent assay was used to determine COX-2 and iNOS levels in the blood serum. At the same time, statistical analysis was performed using Pearson product-moment correlation coefficients, regression, and receiver operating characteristics analyses. Our follow-up study examined the relationship between COX-2 and iNOS in the blood serum of >500 Turkish patients with different types of pain (Sadik, OA, unpublished data, November 2021) and assessed their potential as pain biomarkers. Serum COX-2 and iNOS levels were examined along with the level of pain caused by different types of pain, including lumbar or vertebral; lung; osteoporosis; inflammation; and fatigue, headache, or malaise related to problems. The data (Sadik, OA, unpublished data, November 2021) are now used to develop the current BN concept. For more details on our patient samples, see the Methods section.

The BN will estimate the probabilities that the pain the participant is experiencing takes each of the possible levels: 0, 1, 2, 3, 4, and 5. Figure 6 shows the method for exhibiting these estimated probabilities as histograms. The left example shows the computed probabilities when the evidence is fully self-consistent, all pointing to no pain experienced. The right illustration shows computed probabilities when the evidence is not consistent; the participant reports experiencing pain at level 4 (out of 0-5), although the COX-2 evidence points to no pain (out of 0-5), and the iNOS evidence points to a pain level of 1 (out of 0-3). The probabilities over all levels must sum to one. The reader may notice in Figure 6 that we have modeled some “leakage” of probabilities to pain levels adjacent to the levels of the evidence. These behaviors are embodied in our choices of model parameters (conditional probabilities).

Because there is, as yet, no verifiable sensor for experienced pain, such a system would be capable of implementing some form of expert consensus as to how the available evidence should be combined. The best clinical decision support system can alert the caregiver to the extent that the available evidence is consistent. One possibility is to provide additional evidence that may be valuable in reconciling the situation.

Clearly, a point where important decisions are needed involves mapping actual measured biomarker values into the chosen discrete node levels for biomarkers, such as COX-2 and iNOS. Our current working model, the mapping, is presented in Tables 1 and 2. We hasten to point out that these decisions are provided for illustration purposes.

Evidently, the design of the system output can involve a small amount of creativity. Here, the primary source of guidance will be expert opinions from clinicians knowledgeable in this domain. There may be significant disagreements among pain experts, but it seems reasonable to assume that some considerable consensus may be reached regarding the nodes that should be included. Of course, it is always possible to explore different models for different applications. However, a significant challenge remains regarding the setting of many required internal probability parameters. For this, we propose the approach described in Section 3.1. For a somewhat differing approach, the reader may consult Hill et al [59].

Figure 7 shows a distribution from our patient samples as to how consistent or not is the observed evidence that our BN model uses. We quantified the number of steps (node levels) that differed between the reported pain and biomarker levels. If a patient’s difference was fewer than 3 steps for both biomarkers (an arbitrary threshold for illustration purposes), we called it “consistent” and colored it green. Of the 379 patients with all 3 pieces of evidence, 302 (79.7%) had consistent evidence. This suggests that our goal of using these biomarkers to corroborate reported pain may be reasonable. If one or both biomarkers was “inconsistent” (>2 steps different), we highlighted those patients in yellow. The 77 inconsistent patients fell into 5 groups:

1. A total of 8 patients with high COX2, low reported pain (RP), and low iNOS

2. A total of 31 patients with high RP, high COX2, and 0 iNOS

3. A total of 11 patients with high RP, low COX2, and low iNOS

4. A total of 19 patients with high RP, high iNOS, and low COX2

5. A total of 8 patients with high iNOS, 0 RP, and low COX2

A reasonable next step would be to explore the data from these patients for other evidence that may help explain these observations.

The inconsistent evidence may be attributed to cultural or temporal variability of COX-2 and iNOS (as well as their serum variation and half-life), VAS, and other tools. Tables 1 and 2 assume that each categorized COX-2 and iNOS measurement most likely corresponds to a given pain experience. However, it is necessary to perform “sensitivity analysis” for greater precision. Inconsistencies may also be caused by potential miscalibration in pain management [35-37,61]. Ongoing work now includes supplementary information regarding the time-of-day data the biomarkers were taken and measured. The time-course measurements are being compared with other collected data. Information regarding gender and individual differences in COX-2 and iNOS data on RP is also being recorded. Different causes of inconsistencies could be attributed to memory bias and cognitive effects such as exaggerations or underestimations when reporting pain.

In 2018, the National Institutes of Health launched the Helping to End Addiction Long-term Initiative to stem the national opioid public health crisis [60,62]. One component of the Helping to End Addiction Long-term Initiative is to support biomarker discovery and rigorous validation to accelerate high-quality clinical research on pain [62,63]. Biomarkers are objectively measured and evaluated as indicators of either normal or pathogenic biological processes or responses to therapeutic interventions. Multiple studies have observed significant differences in proinflammatory cytokines (eg, interleukin 6 [IL‐6], tumor necrosis factor α, IL‐8, and IL‐1β) in relation to pain intensity [63-65]. Serum protein levels and mRNA expression of tumor necrosis factor α have been shown to be significantly higher in participants experiencing a greater intensity of chronic pain. Biomarkers may be used alone or in combination to assess the health or disease state of an individual [66,67].

Our group conducted extensive research in this area, identifying COX-2 and iNOS or nitric oxide synthase 2 as good candidates for this purpose [58,60,62,63,65,68-71]. COX, also known as prostaglandin H₂ synthase, is a key bifunctional enzyme in the biosynthetic pathway that leads to the formation of prostanoids, including prostaglandins, prostacyclins, and thromboxanes. COX exists in different isoforms [72,73], COX-1, COX-2, and COX-3. COX-1 is an oxidoreductase enzyme constitutively expressed in many cell types. It is presumed to be responsible for the synthesis of housekeeping prostanoids that are critical for normal physiological functions such as regulating vascular homeostasis, gastric mucosa protection, and renal integrity [74]. COX-3 is a variant of COX-1, which has retained intron-1 during translation and is found in human tissues in a polyadenylated form [75]. It is a selective splicing product of COX-1 mRNA with 633 amino acids with less activity in the production of prostaglandin E₂, and it is mainly found in the hypothalamus, spinal cord, and pituitary choroid plexus. COX-2, on the other hand, is usually undetectable in healthy tissues but is rapidly induced and found to be upregulated in a variety of pathophysiological conditions such as neurological diseases [24], pain [76], inflammation, and cancer infection [13,77,78]. Some studies have indicated that the level of COX-2 at the point of inflammation translates to the degree of inflammation and may thus be used to determine the level of inflammatory pain [18]. Nitric oxide (NO) is a highly reactive free radical and, at the same time, an important signaling molecule involved in different functions [79]. Its inducible form, “iNOS,” is expressed in macrophages and other tissues in response to infection or inflammation, generating large amounts of NO in the blood [24,80]. Increased NO levels have been observed during inflammation and arthritis; therefore, iNOS can be considered a pain biomarker.

In addition to these biomarkers, preliminary results from our laboratory indicated that Contactin-1 (CNTN-1) could also be a promising pain biomarker. This is supported by previous studies that pointed to CNTN-1 as a pain suppressor [81,82] and found antibodies against CNTN-1 in patients with chronic inflammatory demyelinating polyradiculoneuropathy [26]. CNTN-1 levels have been shown to decrease in blood in high-pain states, with convergent evidence in other tissues in human studies for the involvement of pain. Anti–CNTN-1 autoantibodies block or decrease the levels of CNTN-1 in chronic inflammatory demyelinating polyneuropathy [22] and have been considered a bona fide pain marker [81-84]. Moreover, human G-protein subunit gamma 7 plays a strong role in signal transduction with decreased levels in high-pain states (ie, it is a pain suppressor gene with transdiagnostic evidence for involvement in psychiatric disorders) [85,86]. Its expression is decreased by omega-3 fatty acids [87,88]. Microfibril-associated protein 3 provides the most robust empirical evidence as a strong predictor of pain in both men and women. It decreases in expression in the blood during high-pain states [28,84,89,90].

The General Secretary approved the institutional review board of the Manisa State Hospitals Union. The study was conducted at Manisa Merkez Efendi State Hospital, Manisa, Turkey. The participants included in the study were recruited from emergency, internal medicine, gynecology, general surgery, clinical microbiology, chest, urology, and physical therapy clinics. Only participants aged ≥18 years, who consented to participate were included in the study. All participant recruitment and data collection were performed by nurses at the clinics.

Patients were excluded from data analysis if they (1) aged <18 years, (2) did not provide sufficient description during anamnesis to determine their level of pain, and (3) had a blood sample not sufficiently large for analysis (Figure 8).

The survey questions were incorporated into the initial intake and anamnesis questions provided by the nurses. The survey questions are presented in Multimedia Appendix 1. The survey questions included information such as participant’s age, gender, smoking and alcohol habits (Figure 9), chronic disease, long-term medication, surgery history, the reason for and duration of the pain, and pain medication before coming to the hospital.

Following this initial questioning, patients were informed of the study. Informed consent was obtained if the patient agreed to participate in this study. Blood samples were collected via the leftover serum from blood samples taken for routine analysis. During the informed consent process, the participants consented to the use of their leftover serum; no patients were asked to donate blood samples specifically for the study.

The pain level for each participant was classified by the nurse performing anamnesis based on the patient’s responses to the survey questions. Pain level was classified from 0 to 5 as 0=no pain, 1=feeling pain but not disturbing, 2=feeling pain and little disturbing, 3=severe pain and requiring painkiller intake, 4=very severe pain and distraction from working and requiring urgent painkiller administration, and 5=unbearable pain requiring urgent painkiller administration and rest as well as causing anxiety. Each pain level with characteristic conditions was explained to the patients, who were asked how they felt and if they had taken painkillers before they arrived at the hospital.

A sandwich enzyme-linked immunosorbent assay was used to monitor the levels of COX-2 and iNOS in the serum, as reported elsewhere [68].