Artificial intelligence-based prognostic modeling in temporomandibular disorders and chronic orofacial pain: a critical conceptual review

3.1 Prognosis versus diagnosis: a fundamental distinction

Diagnostic classification aims to identify the presence or subtype of a disease at a given time point; prognosis, by contrast, seeks to estimate future outcomes, including pain persistence, functional impairment, treatment response, and risk of chronicity [13]. In TMD and OFP, diagnosis alone rarely dictates management decisions. Patients sharing the same DC/TMD diagnosis can follow markedly different clinical trajectories, driven by individual variation in psychosocial factors (e.g., anxiety, catastrophizing, coping style), biological predisposition, pain processing mechanisms, and behavioral and environmental factors. This variability underscores why diagnosis-centric models are insufficient for guiding individualized care [12, 47].

Currently, AI applications in dentistry are primarily based on static or two-dimensional cross-sectional images and focus largely on diagnostic performance [24, 25]. Performance is commonly measured using metrics such as accuracy, precision, F1-score (a composite metric that balances precision and recall), or the area under the receiver operating characteristic curve (AUC-ROC), which quantifies a model’s ability to discriminate between outcome groups [24, 25]. However, these metrics do not directly address the most clinically relevant questions for patients, such as “Will my current pain become chronic?” or “Will this treatment plan be effective?” Addressing such questions requires dynamic information, including temporal patterns and trends, and modeling changes in patient state over time [13].

3.2 Core data domains for AI-based prognostic modeling in TMD/OFP

An effective AI-based prognostic framework for TMD/OFP must integrate multiple data domains that collectively reflect both disease mechanisms and patient experience. Four core domains are particularly relevant (Table 1, Ref. [7, 9, 10, 16, 47, 48, 49]): (1) clinical and demographic data, (2) psychosocial and behavioral factors, (3) imaging and structural data, and (4) patient-reported outcome measures (PROMs). Each domain provides complementary information that, within a multidimensional framework, enables prognostic modeling capabilities unlikely to be achieved using a single-domain approach. The utilization of these data domains in existing studies is discussed in detail in Sections 4 and 5.

3.3 Temporal dynamics and the need for longitudinal modeling

Unlike diagnostic models, prognostic AI systems must account for time. Pain trajectories in TMD and OFP can evolve, remit, or episodically flare in patterns that are potentially influenced by psychosocial factors and therapeutic interventions [10, 16]. The OPPERA studies demonstrated that biopsychosocial characteristics shift over time and that these longitudinal changes are meaningfully associated with TMD outcomes [10]. Approaches such as survival analysis-augmented ML [50], or deep learning architectures such as recurrent neural networks (RNNs), which are designed to learn from sequential data and capture temporal dependencies [51], are therefore conceptually better suited to prognostic tasks than static cross-sectional classifiers. Practically, RNNs and their variants, including long short-term memory (LSTM) networks [51], are designed to process patient data collected across multiple visits, such as repeated pain scores, functional assessments, and psychosocial measures, sequentially. This allows the model to learn how a patient’s current status relates to prior symptom history, making it well-suited for detecting trends such as remission–relapse cycles or slow deterioration that could be missed by a single time-point assessment. In a complementary approach, survival analysis-augmented ML integrates time-to-event analysis with ML, predicting not only whether a clinical outcome (e.g., transition to chronic pain) is likely, but also when it may occur, while capturing complex nonlinear relationships among risk factors. These techniques align with the clinical reality of TMD/OFP, in which patients often follow irregular trajectories of improvement, plateauing, and relapse rather than a predictable linear course.

Most AI studies in TMD and OFP are based on cross-sectional datasets, which are not consistent with the forward-looking nature of clinical prognosis. These studies train models to classify the current state rather than predict future outcomes, a distinction that is often overlooked in study descriptions but is critical for the validity of prognostic claims.

3.4 Explainability and clinical trust

For prognostic AI to inform clinical decision-making, clinicians must understand why a model predicts poor prognosis or treatment resistance, particularly in pain medicine, where shared decision-making is essential [52]. Explainable AI techniques are therefore highly relevant. Local interpretable model-agnostic explanations (LIME) approximate complex model behavior locally, providing explanations for individual predictions in simple, interpretable terms [53]. SHapley Additive exPlanations (SHAP) assign each input variable a contribution score based on cooperative game theory, quantifying its influence on the prediction [54]. These methods, together with feature importance analysis and other model-agnostic interpretability tools, are integral to clinical translation. Models operating as “black boxes” may achieve strong statistical performance but fail to gain clinical acceptance, particularly in conditions influenced by both psychosocial and biomechanical factors [52, 55]. The integrated biopsychosocial AI prognostic framework proposed in this review, encompassing the data domains, temporal integration, modeling, and clinical output stages described in Sections 3.1–3.4, is illustrated in Fig. 1.

Fig. 1.Integrated biopsychosocial AI prognostic framework for TMD/OFP. The framework presents a four-stage prognostic pipeline: (Stage 1) multimodal data collection across four biopsychosocial domains, namely clinical/demographic, psychosocial/behavioral, structural/imaging, and patient-reported outcome measures; (Stage 2) temporal integration, including longitudinal linkage, trajectory modeling, and data leakage prevention safeguards; (Stage 3) AI/ML model development with explainability and validation, including calibration assessment; and (Stage 4) clinical outputs, including individualized risk stratification, modifiable risk driver identification, management pathway linkage, and shared decision-making. Cross-cutting requirements include TRIPOD + AI, PROBAST, calibration assessment, fairness metrics, and stakeholder engagement. (Figure constructed based on the frameworks and evidence discussed in this review using Microsoft® PowerPoint for Mac, V. 16.107). TMD: temporomandibular disorders; DC/TMD: Diagnostic Criteria for Temporomandibular Disorders; PHQ-9: Patient Health Questionnaire-9; GAD-7: Generalized Anxiety Disorder-7; TMJ: temporomandibular joint; MRI: magnetic resonance imaging; CBCT: cone-beam computed tomography; OHIP: Oral Health Impact Profile; NRS: numeric rating scale; VAS: visual analog scale; QoL: quality of life; AI: artificial intelligence; ML: machine learning; RF: random forest; SHAP: SHapley Additive exPlanations; LIME: Local Interpretable Model-Agnostic Explanations; TRIPOD: Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis; PROBAST: Prediction Model Risk of Bias Assessment Tool.

4.1 Imbalance in diagnostic-prognostic ratios

Although interest in AI is increasing in pain management and dentistry, the evidence base for AI-based prognostic modeling in TMD and chronic OFP remains sparse and is markedly skewed toward diagnostic purposes. Most published AI research employs supervised learning to categorize imaging findings, disease presence, or diagnostic categories at a single time point [24, 26, 27]. Although these studies frequently report high levels of accuracy, their predictive value for clinically relevant outcomes, including pain persistence, progression of functional impairment, and treatment nonresponse, is usually not evaluated [13, 30].

Recent systematic reviews corroborate this diagnostic-prognostic disparity. Jha et al. [26] (2022) found that, across several deep learning models used for the diagnosis of TMD, the reported classification accuracies of AI algorithms ranged from 84% to 99.9%; however, all studies were judged to have a consistently high risk of bias, and the certainty of evidence was therefore rated as very low [26]. Farook et al. [27] (2021) similarly reported that ML solutions in oral and OFP management primarily focus on diagnostic tasks instead of outcome prediction, highlighting a significant gap in the development of prognostic tools for patient care and treatment planning in these fields. Both reviews indicate that the literature largely lacks actual prognostic modeling characterized by temporal validation, longitudinal outcomes, and clinically interpretable endpoints [26, 27]. This distinction is important, as many studies described as “predictive” in the TMD AI literature do not fulfill established criteria for prognostic research and are more accurately classified as diagnostic AI [17, 38, 56].

To summarize the landscape of existing research in relation to the first objective of this review: the types of outcomes predicted have largely been limited to diagnostic classification rather than true prognostic endpoints; the data modalities utilized have been predominantly imaging-based or limited to single-domain clinical variables; and the modeling approaches adopted have mainly relied on supervised learning classifiers applied to cross-sectional data.

4.2 Data from other chronic pain areas

Given the limited number of prognostic AI studies specific to TMD/OFP, evidence from related chronic pain fields may offer useful reference points; however, such insights must be interpreted with careful consideration of the clinical context.

In examining ML techniques in pain research, Lötsch and Ultsch (2018) demonstrated the potential of ML techniques to detect complex patterns in clinical data and to predict pain phenotypes [57]. Parallel analyses in low back pain, conducted by Tagliaferri et al. [58] (2020), found no studies employing AI specifically to predict low back pain prognosis, a gap that directly parallels the situation in TMD/OFP. The evidence reviewed suggests that shared prognostic factors, including psychological components, baseline interference, and pain history, are relevant across TMD and other chronic pain populations [10, 16]. Nevertheless, the presence of shared factors does not guarantee direct transferability. TMD/OFP involves jaw-specific anatomical, behavioral, and psychological elements, including jaw function, parafunctional habits, and associated anxiety and sleep disturbances, that necessitate domain-specific prognostic models [13, 52]. A systematic review and meta-analysis of deep-learning models for TMJ arthropathies further demonstrated high diagnostic sensitivity and specificity, however, the field remains primarily diagnostic rather than prognostic in orientation [59].

4.3 Growing prognostic indicators from studies on TMD

Although a dedicated focus on AI-based prognosis has been lacking (Table 2, Ref. [16, 20, 21, 26, 27, 48, 57, 60, 61, 62, 63]), a few studies have begun to examine prognostic factors within the context of TMD. Most prognostically relevant data have been derived from the OPPERA study. In a prospective cohort of 3263 participants, Fillingim et al. [16] (2013) identified several premorbid psychological characteristics, notably the severity of somatic symptoms, as predictors of first-onset TMD. Bair et al. [48] (2013), using lasso regression and random forests to analyze 202 baseline risk factors from the OPPERA study, found that the most important predictors of incident TMD were comorbid pain, psychosocial parameters, and baseline pain sensitivity [48]. Supervised cluster analysis was used by Bair et al. [60] (2016) to develop a classification system that identified three biopsychosocial subgroups, designated as the adaptive, pain-sensitive, and global symptoms clusters. These clusters were significantly predictive of both the prevalence and incidence of TMD [60]. More recently, Al Turkestani et al. [61] (2024) developed a comprehensive patient-specific prediction model for TMJ osteoarthritis (DC/TMD: defined degenerative joint disease) progression using an ensemble framework that integrated clinical, imaging, and biological variables with clearly defined longitudinal outcomes over a 2–3-year follow-up, further illustrating the feasibility of a prognostic-first AI approach in TMD [61].

Although nearly all studies in Table 2 were rated as having a high risk of bias, the evidence base is not entirely at a standstill. Certain studies represent partial steps toward methodological maturity: Al Turkestani et al. [61] incorporated prospective longitudinal follow-up with multimodal data integration, while the OPPERA-based analyses [16, 48, 60] demonstrated the prognostic value of large cohort, multidimensional data collection approaches. However, no single existing study satisfies all criteria required for a well-developed prognostic AI model in TMD/OFP, namely, a prospective design, multimodal biopsychosocial data, clearly defined longitudinal outcome, external validation, calibration assessment, model explainability, and formal bias evaluation. The field is therefore best described as being at an early foundational stage, with individual studies contributing important building blocks, although none yet provides a complete template for future work.

4.4 Imaging-centered AI models: limited prognostic value

Imaging-based AI models, which have typically been developed to analyze panoramic radiographs or cone-beam computed tomography images for TMJ osteoarthritis diagnosis (DC/TMD: defined degenerative joint disease), are among the most technically developed areas of AI research in the TMD field [26, 46]. Deep learning models have demonstrated strong performance in detecting structural abnormalities in the TMJ [64, 65, 66]. Nevertheless, structural abnormalities in the TMJ have been reported to have a weak or even negligible relationship with patients’ clinical signs and symptoms, including pain intensity, functional disability, and long-term prognosis [47]. Moreover, current imaging-based models do not incorporate functional indicators such as jaw movement dynamics or masticatory muscle activity, nor do they integrate neurological biomarkers such as quantitative sensory testing profiles or indices of central sensitization, which may carry greater prognostic relevance for pain persistence than static structural findings alone.

This limitation is clinically important. Over-reliance on imaging-based prognostic signals may perpetuate outdated biomedical thinking and lead to unnecessary interventions rather than informed, conservative management [5, 6], which can worsen patients’ conditions. Such an approach also neglects the importance of holistic care. More generally, prognostic models that prioritize structural imaging over psychosocial factors may produce inaccurate predictions in relation to chronic pain, which is inconsistent with the goals of patient-centered, conservative care. A further risk that deserves explicit attention is over-medicalization. When AI models are trained predominantly on imaging features, such as condylar morphology, disc position, or joint space measurements, without adequate weighting of psychosocial phenotypes, structural variations that may be incidental rather than pathological can be inappropriately flagged as clinically significant. For example, disc displacement, including DC/TMD disc displacement categories, has been reported in a substantial proportion of asymptomatic individuals [67], and mild condylar remodeling may reflect a physiological process rather than disease progression. An AI model that flags these findings as high risk, without considering the patient’s pain experience, psychological status, or functional impact, does not resolve a clinical issue; instead, it may create one. Such models may lead to unnecessary imaging referrals, invasive procedures, or patient anxiety, all of which are inconsistent with the conservative biopsychosocial approach supported by current evidence for the management of TMD [6, 7, 9, 52]. For prognostic modeling in TMD/OFP, structural data should therefore be contextualized as one input among many within the patient’s broader clinical and psychosocial profile, rather than being used as the dominant or sole basis for risk stratification.

5.1 Study design misalignment

A fundamental limitation across the existing literature is the mismatch between study design and prognostic objectives. Prognostic modeling requires longitudinal data with clearly defined follow-up periods and outcome measures [4, 38]. Many AI studies labeled “predictive” in TMD and OFP have relied on cross-sectional datasets or short-term follow-up, making true prognostic inference methodologically invalid [30]. Consequently, these studies often produce models that predict contemporaneous states rather than future outcomes, a distinction that is frequently obscured in reporting [17, 56]. Given that most reviewed studies lacked a clear temporal separation between input variables and outcomes, the risk of data leakage, as discussed in Section 5, cannot be excluded, further undermining the validity of their prognostic claims.

5.2 Outcome definition and clinical relevance

Existing studies frequently utilized highly variable and poorly justified prognostic endpoints of limited clinical relevance, such as changes in pain thresholds or reductions in imaging abnormalities [14, 39]. These endpoints often bear little relationship to clinically meaningful outcomes, such as overall treatment response, degree of clinical recovery, and risk of persistent clinically significant pain states [6, 7]. Thus, such studies may yield statistically significant prediction models that have limited bearing on clinical decision-making in practice [17].

5.3 Sample size, data quality, and overfitting

Small sample sizes are a recurring limitation in the reviewed studies. ML models, particularly high-dimensional models, are prone to overfitting when trained on limited data [56]. Overfitted models may perform well internally but degrade substantially when applied to new patient populations. Compounding this limitation, many studies failed to report details on data preprocessing, missing data handling, or class imbalance, all of which can substantially influence model performance and reproducibility [30, 69]. These reporting gaps have also been highlighted in a commentary on the reporting of AI prediction models, which identified heterogeneous design and limited reproducibility as recurring concerns across clinical domains [70]. These issues mirror the broader pattern identified by Wynants et al. [68] across COVID-19 prediction models.

5.4 Absence of external validation

The process of validating models on independent datasets is essential to the development of any robust prognostic model [34, 71]. Studies of AI in OFP rarely include external validation, and validation methods beyond cross-validation are seldom used [34, 71]. External validation is particularly important in TMD/OFP due to the large number of relevant variables, including diverse psychosocial risk factors, variable levels of access to healthcare, and differing clinical practice patterns, all of which may affect the performance of a model developed under different conditions. Without external validation, the generalizability of these models remains uncertain [70]. The concept of domain shift is especially relevant to OFP management. Most AI models were developed in tertiary orofacial pain centers, where patients tend to present with more complex, chronic, and treatment-resistant conditions. When these models are applied in primary care or general dental settings, where patients often present with earlier-stage disease and different psychosocial profiles and comorbidities, they may systematically overestimate chronicity risk or misclassify treatment response. This problem is further compounded by differences in clinical assessment protocols (e.g., extent of DC/TMD Axis II evaluation), referral thresholds, socioeconomic context, and cultural variation in pain expression across geographic regions [34, 71]. Domain shift is therefore not merely a statistical issue but also a clinical concern, affecting equitable and appropriate pain management. This underscores the necessity of validating prognostic models across geographically and sociodemographically diverse populations.

5.5 Reporting quality, model transparency, and explainability

Complex and opaque algorithms may pose barriers to practical clinical use, even though deep learning methods can model more complex, nonlinear patterns than simpler algorithms. Such models are generally less interpretable, and clinicians must reconcile this limited interpretability with the need for clinical and social trust, participation in shared decision-making, and adherence to ethical principles [18, 55]. In pain medicine, where clinical estimates of prognosis routinely inform patient education, counseling, and communication of post-surgical expectations, a TMD/OFP model that produces a prognostic estimate should also provide a corresponding clinical explanation. Complex models may give the impression of interpretability that exceeds their actual transparency and may therefore fall short of the clinical utility requirements in practice [55].

As noted above, none of the prognostic studies examined reported model calibration, relying exclusively on discrimination metrics such as AUC and accuracy (Table 2). In prognostic modeling, calibration—the alignment between predicted probabilities and actual outcome frequencies—is often more clinically relevant than discrimination, as it directly affects the reliability of personalized risk estimates communicated to patients and healthcare professionals [28, 29]. When a prognostic model reports strong discrimination without calibration, there is no assurance that a predicted probability (e.g., 70%) corresponds to an actual observed risk near that value, a gap that could lead to inappropriate treatment escalation or false reassurance if acted upon clinically [14, 17].

5.6 Bias and equity

Bias amplification represents a further, often overlooked risk. When AI prognostic models are trained on datasets that inadequately represent specific demographic or psychosocial subgroups, they may generate systematic errors in risk assessment, potentially exacerbating existing inequalities in pain management [33, 35]. This issue is not merely theoretical. Obermeyer et al. [35] demonstrated that a widely used commercial healthcare algorithm systematically underestimated the health needs of Black patients because it relied on healthcare cost as a proxy for illness severity, a measure influenced by access to care rather than true clinical need. This finding is particularly relevant for TMD and OFP, where sex-based prevalence differences, psychosocial factors linked to socioeconomic status, and differential access to specialized care can similarly distort model training and subsequent prognostic predictions [33]. If prognostic AI models for TMD/OFP are developed without attention to these factors, they may inadvertently perpetuate or worsen disparities. Few existing studies assess fairness, subgroup performance, or bias mitigation strategies. A similar concern applies to DC/TMD Axis II measures: high somatic symptom burden scores may reflect true symptom severity but can also be influenced by reporting style, cultural context, or comorbidities. If such scores are used uncritically in model training, the resulting model may treat somatic symptom burden as inherently prognostic of chronicity, thereby reinforcing assumptions rather than empirically testing them. Reliance on such outputs risks generating self-fulfilling prognostic predictions. Psychosocial variables should therefore be interpreted within validated clinical frameworks, and bias auditing should be a standard component of model development.

6.1 The prognosis question nobody is asking

All current AI studies were designed around diagnostic questions, with prognostic claims made only retrospectively. As discussed in Section 5.1, this design misalignment limits the prognostic validity of existing models. To close this gap, future studies should adopt a prognosis-first design, beginning with clinically relevant prognostic questions, such as chronicity risk, likelihood of success with conservative therapy, and risk of symptom recurrence, and should define cohort selection criteria, follow-up duration, outcome definitions, and modeling approaches accordingly [4, 38, 71]. Standardized data collection based on validated assessment tools, such as DC/TMD Axis II and patient-reported outcome measures (PROMs), would facilitate cross-study comparison and multicenter collaboration.

6.2 The cost of ignoring time

As highlighted in Section 5.1, the current literature is predominantly based on cross-sectional data. Consequently, the temporal characteristics of pain trajectories are not captured. Temporal modeling approaches discussed in Section 3.3, including survival analysis-augmented ML [50], mixed-effects modeling combined with ML, and RNN architectures [51], have not yet been applied in the context of TMD/OFP. However, evidence from other pain and medical fields suggests that these modeling approaches could enable a more nuanced understanding of the temporal characteristics of pain, potentially identifying key transition points amenable to early intervention.

6.3 Why are psychosocial data still an afterthought?

As discussed in Section 4.4, current AI models continue to underrepresent psychosocial variables. Without such data in training datasets, algorithms learn to depend on imaging and clinical features. This creates a self-reinforcing cycle whereby the absence of psychosocial data today makes their inclusion in future studies less likely [9, 10, 11, 12, 16, 63]. Imaging data constitute only one component of a comprehensive prediction model, and biopsychosocial integration is essential for meaningful prognostic research. There is also a technological dimension to this problem. ML algorithms tend to assign greater weight to clinical and imaging variables, not because these are inherently stronger predictors, but because they are collected more consistently, carry less measurement noise, and are represented in standardized numeric formats. Psychosocial data, by contrast, are typically gathered through self-report instruments that can exhibit floor and ceiling effects, vary across cultures, and appear less informative to algorithms optimized for structured inputs. Overall, these considerations, together with evidence repeatedly demonstrating psychosocial factors as important prognostic markers [9, 10, 16], underscore the need to break this cycle by purposefully incorporating validated psychosocial assessment tools into future model development.

6.4 Models that have never left the lab

The near-universal absence of external validation, as discussed in Section 5.4, means that no current model can be safely assumed to generalize beyond its development cohort. Multicenter datasets, federated learning [72, 73], and transparent reporting frameworks such as TRIPOD + AI [29] can improve generalizability while protecting data privacy. Federated learning is particularly promising, as it enables multiple institutions to jointly train a shared model without exchanging raw patient data. Prospective validation studies, although resource-intensive, carry particular value for establishing clinical credibility and informing regulatory pathways [34, 71].

6.5 When predictions do not lead to decisions

Many AI models produce probabilistic outputs, which can be difficult to interpret in clinical settings. Prognostic estimates that do not map onto actionable thresholds, such as low-, moderate-, or high-risk categories, offer limited utility to practitioners managing complex pain conditions [6, 7]. Explainability should be prioritized alongside predictive performance. To ensure model outputs align with clinical reasoning and workflow, hybrid modeling approaches that balance interpretability and accuracy should be adopted, incorporating clinician-in-the-loop development processes [18, 55]. In practical terms, a prognostic model for TMD/OFP should not merely yield a probability of chronicity but should translate that estimate into recommendations that clinicians can interpret and that align with established care pathways. For example, it should flag when a patient’s psychosocial risk profile suggests a low chance of responding to conservative splint therapy alone, prompting early consideration of multidisciplinary interventions, including cognitive-behavioral therapy. Importantly, the probability thresholds that trigger transitions between care levels (e.g., from conservative management to multidisciplinary intervention) cannot be prescribed a priori; they must be calibrated through prospective studies that map model-derived risk strata onto observed treatment outcomes across diverse clinical populations. Explainability methods, such as SHAP and LIME, can identify which variables drive individual predictions, but their values are model-specific and should not be interpreted as universal clinical cutoffs. Close collaboration among clinicians, data scientists, and behavioral researchers is essential to ensure models highlight modifiable risk factors, such as stress, sleep disturbance, and parafunctional behaviors, enabling targeted interventions. Fig. 2 illustrates the proposed end-to-end prognostic AI pipeline for TMD/OFP, spanning from multimodal biopsychosocial data collection to prospective outcome evaluation.

Fig. 2.Proposed prognostic AI pipeline for OFP and TMD. The pipeline spans six stages: (1) multimodal baseline data collection, (2) longitudinal data integration, (3) AI/ML model development with explainability and validation, (4) prognostic outputs, (5) clinician-in-the-loop decision support, and (6) prospective outcome evaluation with feedback for model refinement. Cross-cutting requirements include TRIPOD + AI, PROBAST, calibration assessment, fairness metrics, and stakeholder engagement. (Figure constructed based on the frameworks and evidence discussed in this review using Microsoft® PowerPoint for Mac, V. 16.107). AI: artificial intelligence; ML: machine learning; DC/TMD: diagnostic criteria for temporomandibular disorders; PHQ-9: patient health questionnaire-9; GAD-7: generalized anxiety disorder-7; TMJ: temporomandibular joint; MRI: magnetic resonance imaging; CBCT: cone-beam computed tomography; NRS: numeric rating scale; VAS: visual analog scale; QoL: quality of life; OHIP: oral health impact profile; SF-36: 36-item Short Form Health Survey; RNNs: recurrent neural networks; SHAP: SHapley additive explanations; LIME: local interpretable model-agnostic explanations; CV: cross-validation; TRIPOD: Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis; PROBAST: Prediction Model Risk of Bias Assessment Tool.

To translate prediction into decision support, it is necessary to define what constitutes an actionable prognostic output in TMD/OFP care. A binary classification, such as “high risk of chronicity”, or a probability score in isolation, provides limited guidance to the treating clinician. A more useful output should not only estimate risk but also identify which modifiable factors contribute most to that risk for an individual patient. For instance, if a model determines that a patient’s elevated risk of persistent pain is driven primarily by catastrophizing, sleep disturbance, and parafunctional habits rather than by structural joint findings, this information could guide referral toward cognitive-behavioral therapy, sleep hygiene counseling, or behavioral modification, rather than further imaging or surgical consultation. In this way, the model output becomes directly linked to a specific, individualized management pathway. Achieving this level of clinical utility requires that prognostic models incorporate explainability methods [53, 55] and that outputs are developed collaboratively with clinicians, ensuring the information provided is both statistically valid and interpretable within existing care workflows [39, 74]. Without this step, even a well-calibrated and externally validated model is unlikely to meaningfully impact clinical practice.

6.6 Equity and accountability: unfinished business

As discussed in Section 5.6, bias assessment is largely absent from existing studies. Given the well-established sex differences and psychosocial disparities in TMD/OFP prevalence and outcomes [4, 16], explicit assessment of subgroup performance, fairness metrics, and bias mitigation strategies should be regarded as a minimum standard for future research [33, 35]. Table 4 summarizes the identified research gaps, along with suggested actions and their corresponding priority levels. In addition, Fig. 3 presents a phased research roadmap, illustrating the progression of AI-driven prognostic modeling from initial data collection to eventual clinical application.

Fig. 3.Proposed research roadmap for AI-based prognostic modeling in TMD/OFP. This figure presents a staged research roadmap spanning three phases: (Phase 1) Foundation, establishing standardized longitudinal data collection with biopsychosocial integration and validated outcome measures; (Phase 2) Model Development, applying trajectory-based ML methods, integrating explainability, and conducting rigorous internal validation; and (Phase 3) Translation, involving multicenter external validation, federated learning implementation, clinical decision support integration, and prospective outcome evaluation. Black arrows indicate progression from one phase to the next. Each phase includes recommended methodological standards (TRIPOD + AI, PROBAST), ethical safeguards, and milestones for assessing readiness to proceed (shown using colored arrows). (Figure constructed based on the frameworks and evidence discussed in this review using Microsoft® PowerPoint for Mac, V. 16.107). DC/TMD: Diagnostic Criteria for Temporomandibular Disorders; ML: machine learning; PROMs: patient-reported outcome measures; RNNs: recurrent neural networks; SHAP: SHapley Additive exPlanations; LIME: Local Interpretable Model-Agnostic Explanations; k-fold CV: k-fold cross-validation; TRIPOD + AI: Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis plus Artificial Intelligence; EHR: electronic health record; PROBAST: Prediction Model Risk of Bias Assessment Tool.

7.1 Communicating uncertainty without doing harm

Unlike diagnostic classifications, the value of a prognostic prediction lies primarily in its impact on clinical management. The uncertainty inherent in prognostic predictions can affect the management of chronic pain conditions, where patients’ beliefs and expectations play a crucial role in their clinical outcomes. Providing a patient with a probabilistic risk estimate for developing chronic pain can influence their symptom experience, coping behavior, and adherence to treatment recommendations [10, 11]. A binary prediction indicating that a patient is at high risk of developing chronic pain can reinforce maladaptive beliefs and catastrophizing in psychologically vulnerable patients [9, 16].

Communicating uncertainty in prognostic prediction can therefore complicate chronic pain management. Development of AI-based tools to guide communication of risk estimates could support shared decision-making with patients. As with diagnostic predictions, prognostic AI tools must communicate predictions as risk estimates with confidence ranges rather than as definitive outcomes, in order to avoid undermining the patient’s ability to manage their condition and focus on modifiable risk factors such as stress management, sleep hygiene, and adherence to conservative therapy [10, 11].

7.2 Explainability as an ethical necessity in prognostic AI

Explainability is not merely a technical preference but an ethical necessity in prognostic AI applications. Clinicians must understand which variables contribute most strongly to prognostic estimates to justify clinical decisions and identify potentially modifiable risk factors [18, 55]. Methods such as LIME [53] and SHAP [54] can substantially enhance interpretability without sacrificing predictive performance, enabling responsible use in clinical contexts.

7.3 Whose pain counts? Bias and fairness in prognostic AI

Ethical model development requires explicit evaluation of subgroup performance, transparency regarding data sources, and the application of fairness metrics [33, 35]. A model that exacerbates disparities in pain management, either by undertreating higher-risk subgroups or overtreating lower-risk subgroups, is ethically unacceptable. As Char et al. [75] argued, general ethical considerations for the use of ML in healthcare—such as informed consent, data privacy, algorithmic transparency, accountability for clinical decisions, and potential for bias—are broadly applicable but carry particular importance in pain management, given the central role of subjective experience, individual preferences, and social determinants of health [75].

7.4 Keeping the clinician in the loop, and making it work in practice (clinical responsibility)

AI-based prognostic models should function as decision-support tools, not autonomous decision-makers. The treating clinician should retain full clinical responsibility and use model predictions in combination with their own assessment, clinical judgment, patient preferences, and clinical context [6, 7]. Price et al. [76] have highlighted the need for legal frameworks addressing liability when AI tools inform clinical decisions [76]. Such tools must always be interpreted as decision aids rather than autonomous systems [74].

For practitioners managing TMD and chronic OFP, AI-based prognostic modeling offers the potential to improve risk stratification, patient communication, and treatment planning (Table 5). Even after an exact DC/TMD diagnosis, patients with the same diagnosis are often managed differently across clinics and over time, and the condition can progress at markedly different rates [1, 5]. Prognostic models integrating clinical, psychosocial, and behavioral data could help identify patients at higher risk of persistent pain or functional deterioration early in the care pathway, enabling tailored management, including closer monitoring, earlier multidisciplinary involvement, or targeted psychosocial interventions for high-risk individuals, while avoiding overtreatment in lower-risk patients [6, 7].

These strategies must be implemented without imposing undue burden on practitioners [75, 76]. Ideally, tools would be integrated within electronic health record systems or embedded in existing workflows where data are already collected during a DC/TMD Axis I and II evaluation. Clinician involvement in tool design is essential to ensure usability, relevance, and trust. AI outputs should always be interpreted within the context of a comprehensive clinical assessment, and the resulting recommendations should not be considered mandatory or definitive recommendations. Natural language processing (NLP) offers a potential approach to incorporating psychosocial information into prognostic models. Clinical notes in electronic health records frequently contain rich narrative accounts of patients’ pain experiences, coping strategies, sleep patterns, and social contexts, which are routinely recorded by clinicians but seldom used for model training. NLP methods could extract and structure this narrative data, potentially providing a more clinically meaningful representation of a patient’s psychosocial profile than questionnaire scores alone. Although NLP applications in TMD/OFP have yet to be explored, their increasing use in other chronic pain domains suggests that this may represent a promising direction for future research [77].