Thoracic (T8–T10) contusion, compression, or transection models induce post-injury ectopic activity, abnormal spinal and thalamo-cortical rhythms, and glial responses that together sustain neuropathic pain (80). Induction requires microsurgical exposure and standardized impactors (weight-drop or electronically controlled) to deliver reproducible parenchymal injury, followed by neurological and pain assessments (81). In 2017, the Louisville Injury System Apparatus (LISA) introduced displacement-controlled contusion in mice using laser range sensing plus software control, achieving highly consistent T10 injuries (82). Fully automated mouse contusion systems subsequently combined laser distance sensing with motorized stages and intelligent software to produce user-defined SCI severities with improved impact localization, parameter standardization, and operational ease (83). Finite-element (FE) models now map injury mechanics and parameter sensitivity, supporting device standardization and cross-platform comparability and enabling prediction of tissue stress/strain during trauma (84).

These models use “loose” ligation to induce partial ischemia and/or demyelination of the target nerve, yielding both stimulus-evoked hypersensitivity and spontaneous pain. For SNL, tight unilateral ligation of L5 (±L6) spinal nerves in rats produces guarding/defensive behaviors on the injured side with concomitant neuroinflammation and glial activation (85). Pain severity is typically indexed by postoperative mechanical and thermal hypersensitivity together with injured-side spontaneous behaviors (86). By ∼4 weeks, the central amygdala shows enhanced astrocytic reactivity and altered neuronal excitability—implicating not only dorsal horn plasticity but also remodeling of ascending–limbic circuits (87); accordingly, facial grimacing and affect-related behaviors are increasingly used as endpoints for spontaneous pain.

BCP features sustained pain as a defining characteristic. Intra-medullary injection of tumor cells into the femur or tibia produces stable osteolysis. Enrichment of inflammatory mediators and NGF drives peripheral sensitization with ectopic discharges, tumor-induced sensory sprouting, and neuroma-like remodeling (90). Within the dorsal horn, widespread central sensitization and heightened synaptic excitability convert persistent peripheral input into tonic pain output even at rest (91). Together, these peripheral and central alterations provide the anatomical and electrophysiological substrate for the spontaneous, unpredictable pain typical of BCP.

Monogenic pain syndromes are modeled by knocking in/out human orthologs with defined variants: SCN9A/Nav1.7 gain-of-function (GOF) causing erythromelalgia, loss-of-function (LOF) causing congenital insensitivity to pain; TRPA1 GOF causing familial episodic pain syndrome; and mutations in NGF/TrkA producing hereditary sensory and autonomic neuropathies (HSAN). These models align closely with clinical phenotypes and thus offer strong translational value. Nonetheless, strain, sex, and age exert substantial influence. Pronounced sex differences have been reported in the molecular and cellular mechanisms of pain (92), implying that genotype effects may be sex-contingent. Both clinical and animal data indicate that aging reshapes the linkage between nociception and inflammation (93). Cross-strain resources and databases can help select the most appropriate strain/age/sex for a given design; multi-background genetic panels have been shown to quantify complex-trait variance systematically and improve reproducibility (94).

Across the above models, common limitations persist: batch-to-batch variability arising from surgical trauma and operator technique; heavy reliance on evoked endpoints with poor sensitivity to spontaneous pain; and observer-dependent, low-throughput scoring (Table 2). As early as 2011, investigators noted that conventional evoked assays overlook spontaneous pain, a core symptom in patients with spinal cord injury (SCI) (95). More recent work linking circuit dysfunction to pain has identified sleep and respiration as behavioral indicators after SCI (96). For example, post-injury spontaneous nociceptor discharges can fragment NREM sleep, whereas analgesic interventions restore sleep continuity (97). These findings indicate that spontaneous pain perturbs endogenous physiological rhythms, underscoring their cross-species promise as physiological readouts. Quantification of such spontaneous behaviors can be achieved noninvasively with home-cage monitoring systems; adding computer vision improves objectivity under unrestrained conditions (98), and integrating machine learning enables more efficient pain detection and grading (99).

Under the combined influence of peripheral inflammation and central sensitization, spontaneous pain can manifest as anatomically specific gait alterations (118, 119). Such functional changes more closely resemble the disability component of chronic pain, because shifts in weight bearing, stance phase, and interlimb coordination can reflect ongoing discomfort rather than a brief withdrawal reflex. Gait analysis is also well suited for evaluating candidate therapies, as sedative or motor suppressive drugs typically do not restore normal walking and therefore are less likely to yield false positive analgesic signals; accordingly, preclinical gait endpoints have shown good reproducibility across studies (120). However, precisely because gait integrates both peripheral drive and supraspinal control, it is vulnerable to confounding by motor impairment, sedation, anxiety, and compensatory strategies, which may not align between rodent models and human patients. Recent NEP focused systematic reviews and syntheses of pain suppressed behaviors therefore recommend interpreting gait as “function under pain” and combining it with measures of comorbidity and affective motivational domains to strengthen translational inference (62).

Common approaches quantify paw-print morphology (e.g., print area/intensity), spatiotemporal parameters (stride length, stance phase), and related metrics to infer pain intensity and quality (121, 122), and are widely used in spontaneous-pain research. The main limitation is limited specificity: gait alterations can be confounded by non-pain factors such as changes in locomotor pattern or environmental stress (123). Another frequently used method scores hind-paw use/weight-bearing after walking—most common in bone-cancer pain and acute inflammatory models—typically with a four-point scale for the affected limb (124). This approach requires minimal equipment and is easy to implement but is highly observer-dependent and time-consuming, limiting its utility in large, high-precision studies. To address these gaps, Krug et al. developed advanced dynamic weight bearing (ADWB), which leverages software innovations to quantify pain severity and has shown greater sensitivity and reproducibility in intra-articular therapy and inflammatory-pain assessments (125). Dent et al. also used ADWB in mice to evaluate ultra-acute hypersensitivity induced by allyl isothiocyanate (AITC); objective, non-reflexive quantification of “loading/unloading” improved measurement objectivity (126). A related continuous hindlimb-use scoring paradigm is commonly applied in peripheral neuropathic models such as sciatic nerve ligation; it performs well for chronic pain but is less sensitive to acute nociception (127).

By comparison, the CatWalk system combines automated video capture with computer-assisted analysis to detect subtler gait perturbations and has been widely adopted across central and peripheral disease models (119, 128). It provides real-time, condition-specific quantification of multiple gait parameters (129, 130). Advantages over traditional methods include: (1) integrated multidimensional metrics—symmetry, stride length, cadence, velocity—which substantially increase resolution for pain-related phenotypes (122); and (2) objective quantification of spontaneous pain and hyperalgesia, improving interpretability and reproducibility (131, 132). Nonetheless, consistency can be suboptimal in certain contexts; for instance, sensitivity is limited in mild mono-arthritis (133). Using well-matched controls (e.g., body weight/length) can mitigate this issue. Cross-platform comparisons are also complicated because systems such as CatWalk, DigiGait, and TreadScan extract non-identical parameter sets.

Effectiveness also varies by rodent strain. Static weight-bearing is highly informative for hind-paw pain in rats but can be challenging in mice. Strain- and sex-specific differences are well documented (133, 134); for example, in C57BL/6 mice, females exhibit more spontaneous pain behaviors than males in chronic degenerative joint pain models (135). So controls should be matched by strain and, when feasible, strain specific reference ranges should be used when interpreting pain related gait changes. All gait-based techniques require standardized behavioral training prior to data collection. Acclimation to handlers and the test environment can markedly influence pain-related performance. Commercial gait platforms such as CatWalk typically require greater equipment investment than direct video recording (119), but they provide automated extraction of multidimensional gait parameters (136), thereby facilitating standardized measurement and higher throughput. In practice, individual acquisition runs are relatively brief, whereas the primary time burden is front loaded to pretest habituation and training and to obtaining a sufficient number of valid trials per animal. Consequently, gait based approaches generally represent a high equipment cost but comparatively low manual scoring burden (137).

Similarly, grimace based pain scoring can capture animals' natural pain responses in the absence of external stimulation. By indexing a spontaneous, affect laden expression of pain, it aligns more closely with clinical pain constructs that rely on central integration, and thus provides a cross species compatible readout of spontaneous pain (138). These methods classify and score pain-linked facial action units. The grimace scales were introduced in rodents over a decade ago and have since been adapted to at least ten mammalian species (139–141).

The Rat Grimace Scale (RGS) (142) and Mouse Grimace Scale (MGS) (143) quantify facial pain by rating parameters such as eye fissure (orbital) narrowing, ear position, nose/cheek bulge, and whisker orientation. These tools exhibit high sensitivity and good test–retest reliability across postoperative, neuropathic, and visceral pain models and are notably effective for detecting the effects of low-potency opioids (144–146). Compared with evoked assays, grimace scales avoid additional injury, suit designs that must minimize intervention, require only a camera and frame-capture software, and demand no animal training—reducing non-pain confounds.

Limitations remain. In postoperative (147, 148) and inflammatory models (149), RGS/MGS are discriminative, yet facial changes often resolve within hours to ∼48 h, lagging the underlying pain course and constraining chronic tracking (150) Some individuals with chronic pain—preclinically and clinically—show minimal overt facial discomfort (151, 152). Specificity can also be context-dependent: RGS tracks dynamic pain changes in colitis (70) but is less responsive in chemotherapy-induced mucositis (146). As with other spontaneous pain measures, grimace based scoring shows both sex differences, often with stronger effects in males than females, and strain dependent variability. For example, one study reported a graded pattern of scores among male mice, with C57BL/6 lower than CD 1 and C3H/He, and a similar strain related gradient was also observed in females (153). In addition, grimace scoring typically relies on manual observation, making it time consuming and susceptible to observer related subjectivity. Reviews of rodent grimace scales indicate that this approach is useful for identifying ongoing pain states and for monitoring analgesic effects, while also emphasizing its sensitivity to contextual factors and methodological choices (154).

To address these issues, computer vision and AI are increasingly integrated into pain assessment. Automated scoring and machine-learning platforms promote inter-lab standardization while enabling parallel appraisal of affective states. Tuttle developed an automated MGS (aMGS) using a deep neural network (InceptionV3), achieving accuracy comparable to human raters (155). Similarly, Arnold et al. reported an automated RGS with ∼97% precision for action-unit detection and reliable prediction of composite RGS scores (156). Parallel advances in human pain assessment include deep-learning systems for postoperative/clinical videos that can partially distinguish genuine from posed pain expressions (157, 158); Kim and colleagues combined Transformer architectures with dynamic video encoding to improve temporal resolution in facial pain tracking (159). Despite continued progress, acquiring, curating, and analyzing high quality facial imagery remains labor intensive, which limits scalability for rapid turnaround and large scale screening. Overall, the method has a low equipment barrier and short per session acquisition time, but its primary bottleneck is the time and personnel required for manual scoring. Automated approaches generally shift effort upstream, requiring greater initial investment in data preparation and model development, while substantially reducing downstream scoring and analysis workload (160). However, grimace remains an indirect readout of endogenous inhibition and is best paired with assays that explicitly probe descending control or pain relief–driven preference, such as CPP or CPA.

Regarding reliability, systematic reviews indicate strong inter- and intra-rater agreement in rodent models, with weaker evidence in other species (e.g., rabbits, pigs) (150). These constraints have accelerated multimodal strategies—for example, combining noninvasive behavioral indices (nesting, burrowing) with grimace scales (161), or synchronizing facial metrics with neuroimaging, electrophysiology, and thermoregulatory readouts to establish convergent validity (162, 163). Emphasizing noninvasive techniques reduces stress and aligns with the 3Rs in the development of pain-assessment systems (156).

In recent years, conditioned place preference/aversion (CPP/CPA) has been used increasingly to interrogate spontaneous pain in rodents. The core logic is to measure preference or avoidance shaped by tight pain–context associations: during ongoing pain, an intervention that relieves pain becomes associated with a specific context and yields CPP; conversely, contexts paired with pain are avoided (CPA). CPP/CPA directly probes the affective and motivational dimensions of pain, and this conceptual framework has been validated in systems neuroscience. Mechanistically, pain aversion and the rewarding value of pain relief are encoded by midbrain–limbic circuitry. Accordingly, CPP/CPA can quantify otherwise hard-to-observe ongoing spontaneous pain rather than mere changes in reflexive responses, thereby better approximating the true disease burden of chronic pain (164, 165). The apparatus is relatively standardized and does not require expensive imaging systems, but it does require compartmentalized chambers, careful control of contextual cues, and behavioral tracking. Because much of the procedure remains manual, the time burden is substantial, typically involving multiple days of habituation, conditioning sessions, and testing (166).

In practice, CPP is frequently deployed for mechanism validation under precise interventions. For example, selective blockade of afferent transmission from the injured segment by dorsal root ganglion field stimulation (GFS) reduces spontaneous dorsal horn firing in nerve-injury models and produces analgesia-induced CPP, thereby closing the causal chain of “spontaneous pain → dorsal horn hyperexcitability → reward-like behavior” (167). Inhibiting T-type Ca²⁺ channels (Cav3.2) in peripheral sensory neurons alleviates both spontaneous and evoked pain (168). CPP also gauges abuse/reward liability by testing whether a candidate analgesic elicits CPP in the absence of pain relief, separating true analgesia from drug seeking (169). Recent studies further report sex differences in CPP readouts in some trigeminal neuralgia models, underscoring the need to incorporate sex in design and interpretation (170). Whereas CPP emphasizes negative reinforcement within VTA/NAc circuits, CPA engages cortical–limbic pathways (e.g., ACC) and does not rely on relief-reward as an indirect proxy; it is thus well suited to visceral-pain affect. In bone cancer pain, manipulating TrkB, NR2B, or ERK–CREB selectively weakens avoidance without altering mechanical thresholds, indicating that CPA maps more closely onto pain's affective dimension (20). Meta-analytic summaries of CPA emerging 1–2 weeks after inflammatory insult show time sensitivity and reproducibility for the affective facet of spontaneous pain (171).

CPP and CPA capture the affective and motivational dimension of pain and can be linked to circuit level readouts, which supports their cross species potential and translational relevance. In the context of endogenous pain modulation, descending inhibition or facilitation can shift the net aversiveness of ongoing pain and the value of analgesic relief, thereby changing preference or avoidance strength. This makes CPP and CPA, to some extent, a functional readout of central regulation rather than a mere surrogate of peripheral sensitivity, and such central modulatory differences are often more pronounced in chronic pain states. Consistent with this view, conditioned pain modulation in humans is commonly considered the behavioral counterpart of diffuse noxious inhibitory controls in rodents and is used to probe descending inhibitory function (172). Cross species credibility is further strengthened by circuit evidence. In mice, expectation and higher order appraisal can recruit specific cortical pathways to produce analgesia, a core feature of central modulation also emphasized in human studies (173). Human molecular imaging and fMRI work on the nucleus accumbens and related reward circuitry likewise supports translational mapping between pain aversion and relief related reward (174, 175). In clinical practice, CPM is typically operationalized using dynamic QST paradigms, in which changes in the perceived intensity of a test stimulus during a spatially remote conditioning stimulus are used to index descending inhibitory function. By contrast, DNIC in animals is often assessed under anesthesia or electrophysiological settings, with suppression of dorsal horn neuronal activity or spinal reflex outputs as the primary readout (176). Notably, although both paradigms capture “pain inhibits pain” related descending modulation, CPM is more susceptible to higher order influences such as attention, expectancy, and affect, and therefore is not fully equivalent to DNIC (172). Moreover, contextual preference forms only when animals are in an aversive, spontaneous-pain state and analgesia is paired with a specific context, conferring greater specificity relative to short, acute evoked pain. Finally, CPP and CPA can be coupled with electrophysiology, in vivo imaging, and circuit manipulations to strengthen mechanistic inference. For example, lactate signaling in anterior cingulate cortex astrocytes has been detected during the formation and retrieval of visceral pain CPA (177) and fiber photometry and single unit recordings have reported VTA, nucleus accumbens, and anterior cingulate cortex activity aligned to CPP or CPA expression (178). Together, these features make CPP and CPA well suited for testing interventions that target central modulation by asking whether they truly reduce the aversiveness of ongoing pain and the value assigned to analgesic relief, supported by convergent multimodal evidence.

As with other assays, CPP/CPA is influenced by sex, strain, and stress (179). Classic methodological reviews also note confounds from individual differences in learning/motor ability and from apparatus variability (180). In line with recent trends, these limitations can be mitigated by: (1) high-throughput video co-acquisition of spontaneous behavior to integrate CPP/CPA with markerless pose estimation and interpretable pain-phenotype extraction (181); (2) synchronizing CPP/CPA with in vivo calcium imaging/electrophysiology to record circuit activity in real time (e.g., fiber photometry of ACC projections and NAc dopamine) (165); and (3) ML-based feature extraction with individualized thresholds to unify high-throughput video, calcium, and electrophysiology, thereby reducing context/motor confounds and improving inter-lab reproducibility and translational value.

For evoked assays such as von Frey, Hargreaves, and hot-plate tests, the key bottlenecks are bias during stimulus delivery/observation and low throughput. Dedek designed an integrated, computer-controlled platform for thermal and mechanical stimulation and trained a neural network for automated measurement—while also initiating video-based capture of spontaneous behavior (75). Pairing high-frame-rate video with machine learning enables millisecond-scale, unbiased kinematic analysis of paw withdrawal (182). Three-dimensional, multi-camera capture with keypoint tracking mitigates the “reaction-only, context-blind” limitation of evoked tests. Although generic clustering may mistake high-frequency jitter for target actions, a keypoint-MoSeq pipeline can discover behavior modules from keypoints without supervision (e.g., rapid sniffing) (183). In diabetic neuropathy (DN) models, 3D behavior analysis combined with unsupervised learning (MoSeq) yields finer phenotyping and treatment sensitivity (184); MoSeq likewise strengthens gait analysis by reducing dependence on running speed and training, improving cross-lab reproducibility. Automated micro-movement detection further reduces observer bias: the Facemap framework couples keypoint tracking with a deep encoder to extract eyelid aperture, whisking, and nose–mouth micromotions from video, delivering high-throughput grimace analysis while predicting neural activity (185). Relatedly, DeepLabCut-based pipelines annotate eyelid distance/palpebral fissure to quantify mouse facial pain (186). Butler and colleagues fluorescently labeled distal limbs and used multi-view, variable-illumination video to auto-generate large training sets, improving generalization to fine movements—a strategy that may remedy weak whisker-unit detection in grimace scales (187). These modern approaches entail substantial upfront costs, including specialized hardware, computational resources, data annotation, and model training. Once the pipeline is established and validated, however, the per animal manual workload drops markedly, enabling greater scalability (188).

It is important to note that automation and higher apparent accuracy primarily improve measurement consistency and do not necessarily yield more clinically translatable conclusions or deeper mechanistic insight; their value still depends on external validation and systematic control of confounding factors. To mitigate these risks, the development of modern assessment pipelines should prioritize interpretable pain relevant readouts, such as the burden of ongoing spontaneous pain and its affective and motivational dimensions, while concurrently recording potential confounders including locomotor capacity, sedation, and stress to enable separation during analysis. Models should then be externally validated and calibrated across strains, laboratories, and acquisition platforms, with their intended scope of use clearly defined. Reporting can follow frameworks such as TRIPOD plus AI (189), and risk of bias and applicability can be assessed using tools such as PROBAST plus AI (190). Finally, converging behavioral AI readouts with central network endpoints, for example fMRI, or circuit level measures can help distinguish phenotype level association from mechanistic concordance, thereby turning more objective measurement into more reliable translational inference.

Functional MRI offers a clear advantage for discussing pain processing and its translational relevance. It can generate comparable readouts of distributed brain networks in both humans and rodents, allowing endpoints at the level of central processing rather than limiting inference to behavioral reflexes. In a landmark PNAS study, Hess et al. (2011) showed that neutralizing TNF α markedly reduced nociception related activity in the thalamus, somatosensory cortex, and limbic regions within 24 h, indicating that neuroimaging can directly link a molecular intervention to brain level pain processing and providing evidence for central mechanisms underlying rapid analgesic effects (191). More recent rodent fMRI work further demonstrates that distinct pain conditions can produce dissociable patterns of central processing with translational implications. For example, Kreitz, Pradier, Segelcke, combined fMRI with network analysis and classification in incisional and inflammatory pain models to identify condition specific hypersensitivity related network signatures (192, 193). Their results suggest that pain states involve not only changes in sensory pathways but also coordinated alterations in aversion related processing and components linked to descending modulation. The authors further proposed that such signatures could support disease relevant brain biomarker discovery and drug target engagement evaluation, thereby improving the testability of translation.

Building on this framework, incorporating additional modalities can further enhance automated assessment and provide a more complete evidence chain for translational validation of central mechanisms. Ultrasonic vocalizations (USV) index affective state (low-frequency, aversive; high-frequency, positive); combining USV with deep learning may reduce confounds from individual differences in motor capacity (194). Infrared thermography (IRT) non-contactly tracks skin temperature and microcirculation and is widely used to monitor inflammation (195); ML-based denoising, super-resolution, and multi-feature fusion can incorporate such physiological changes into holistic pain scores (196). On the device side, piezoelectric sensing was used as early as 2007 to analyze respiration during sleep with automated pattern recognition (197). Recent open-platform systems built on ultra-sensitive piezos detect the slightest movements—even heartbeats—and integrate video with time–frequency decomposition, clustering, and machine-learning analytics to provide noninvasive behavioral assessment (198). Inevitably, increasing technical complexity introduces new challenges in standardization, data integration, ground-truth labeling, and scalability that must be addressed for robust translation. Combining multimodal measures with central network readouts such as fMRI, and benchmarking them under standardized protocols with consistent metadata capture and external validation, will help multimodal AI pipelines mature into reusable and generalizable pain assessment systems.