From Video To Data: 3D AI Tracking For NHP Behavior

Behavioral endpoints carry real weight in non-human primate (NHP) drug development, especially when a therapeutic effect doesn't show up in biochemical or structural readouts. Locomotion, posture, gait, spontaneous activity, scratching, facial responses: shifts in any of these can be functional evidence of disease progression or treatment response.

The problem is that NHP behavior is continuous and highly multidimensional. Conventional assessment leans on manual observation, video review, or predefined scoring systems. Those work, but they eat time and carry inter-observer variability, which bites hardest when you're trying to quantify subtle or slow-developing changes.

Computer vision and AI offer another route: turn ordinary multi-view video into structured three-dimensional behavioral data. Instead of treating video as a record someone has to sit through, markerless tracking converts movement into quantitative kinematic and ethological endpoints.

So the question for NHP translational work isn't whether AI can "track an animal." It's whether video-derived measurements become scientifically meaningful endpoints for characterizing disease and evaluating pharmacology.

Markerless 3D behavioral analysis pairs synchronized multi-view video acquisition with computer vision and deep-learning-based pose estimation.

No reflective markers, no sensors attached to the animal. The system identifies anatomical key points directly from the video frames, and multiple camera views fill in each other's blind spots, easing the occlusion and perspective problems that plague single-view 2D analysis. From that, you can reconstruct the animal's posture and movement in three dimensions. Prisys' NHP behavioral analysis workflow tracks more than 21 anatomical key points, enough to characterize whole-body movement and joint-related kinematics quantitatively.

The pipeline runs: Multi-view video → 3D pose reconstruction → behavioral segmentation → parameter extraction → pharmacological endpoint

One thing worth being clear about: pose estimation isn't the endpoint. Its value is in turning thousands or millions of individual movement observations into behavioral features you can actually interpret.

Once you have 3D pose data, a whole layer of parameters opens up.

Kinematics first: position, velocity, movement trajectory, posture, gait characteristics, limb coordination, joint-related movement patterns. Then behavior: movement sequences get segmented into recognizable behavioral states and analyzed over defined observation windows. Prisys' platform documentation describes automated decomposition of behavioral sequences with temporal precision down to the scale of seconds, which moves you past isolated behavioral scores toward continuous behavioral profiles.

Instead of writing down that an animal "showed increased activity," you can pin down when activity rose, how long it held, which movement patterns drove the change, and whether the timing lines up with dosing. That's a much finer-grained picture of phenotype and treatment response.

The obvious advantage of markerless tracking: you collect behavioral data without attaching anything to the animal.

Marker-based motion capture gives you detailed movement information, but attaching markers or sensors means extra handling, and the hardware can interfere with the very movement you're trying to measure. Markerless sidesteps that, which matters most when spontaneous behavior is itself the endpoint of interest. Prisys' technical framework combines markerless tracking with synchronized multi-view acquisition and deep-learning-based analysis, automating key-point tracking and behavioral analysis while trimming occlusion and perspective limits.

Markerless doesn't mean validation-free. AI-derived behavioral endpoints still need model-specific validation, and the webinar recap made this point directly: algorithmic outputs should be checked against suitable human-reviewed references and read in the context of the animal model, the study design, and the research question. That goes double when behavioral classification serves as a pharmacodynamic endpoint rather than exploratory observation.

The translational payoff comes when you tie movement data to a specific biological or pharmacological question.

In an efficacy study, "movement" isn't the endpoint. The endpoint is whether treatment reduces a pathological behavior, restores motor function, or changes the trajectory of disease progression.

Three application areas from the Prisys–Bayone NHP BehaviorAtlas webinar show what that looks like.

Scratching is close to a perfect case for objective quantification.

Pruritus studies need scratching frequency, duration, temporal distribution, and movement characteristics. Manual scoring works but gets punishing when observations repeat over hours or days. 3D tracking characterizes scratching as a structured movement sequence rather than a count of visible events, and trajectory, posture, and spatial relationships help separate scratching from grooming or routine limb movement.

At Prisys, this analysis slots into NHP pruritus studies, including the IL-31-induced cynomolgus macaque model, which produces measurable scratching behavior and inflammatory responses for evaluating anti-pruritic and related therapeutic strategies. The behavioral measurements sit alongside clinical observations, inflammatory biomarkers, and pathological assessments as complementary pharmacodynamic endpoints.

Motor phenotypes are the other natural fit.

In NHP models of Parkinson's disease, stroke, and other neurological disorders, the functional changes show up in gait, posture, spontaneous activity, limb use, coordination, or movement velocity. A single observational score struggles to capture all of that. 3D tracking represents movement as a time-dependent dataset, so you can compare behavioral features longitudinally before and after disease induction or treatment. Gait symmetry, movement velocity, posture transitions, and limb coordination get evaluated together instead of as isolated notes.

That matters most in efficacy studies where functional recovery builds progressively rather than arriving as a clean yes/no.

Phenotyping doesn't have to stop at whole-body movement.

The webinar covered 3D facial modeling and expression analysis as a way to catch subtle responses associated with pain or discomfort, quantifying changes in facial geometry and movement instead of leaning on subjective visual interpretation. This is still an emerging application. Read it alongside established clinical observations, physiological measurements, and other study endpoints, and put in the annotation, model-specific characterization, and validation before facial features serve as a formal pharmacodynamic endpoint.

Conventional assessment samples behavior at fixed time points, and that's a real limitation. A treatment-related behavioral change that happens between observation windows goes undetected. Continuous or repeated automated monitoring fills in a denser temporal record.

That matters in pharmacology, where the timing of a behavioral effect is often tied to drug exposure: Drug administration → exposure → pharmacodynamic response → behavioral change

Collect behavioral measurements alongside PK/PD data and you can ask whether behavior changed inside the expected pharmacological window, and whether the size of the response tracks with exposure. The behavioral endpoint becomes one component of a broader translational dataset instead of an isolated observation.

Behavioral analysis earns its keep when it runs alongside the other modalities.

At Prisys, phenotyping folds into broader NHP studies with disease models, PK/PD assessment, in vivo imaging, pathology, and biomarker analysis. It isn't there to replace established endpoints. It adds a functional layer that connects molecular and physiological change to observable phenotype.

The workflow: Behavior → PK/PD → imaging → pathology → biomarkers

Each piece answers a different question. Behavior gives function, PK/PD gives exposure and pharmacological response, imaging gives longitudinal anatomical or functional information, pathology gives tissue-level effects, biomarkers give underlying mechanism. The combination counts most when a treatment produces subtle functional change that no single molecular or structural endpoint fully reflects.

The bigger shift here: video stops being an observational record and becomes a quantitative data source.

Hours of footage hold hours of behavioral information, but pulling it out by hand is slow and inconsistent. Markerless 3D tracking turns the same footage into structured datasets containing pose, movement, behavioral sequences, and temporal patterns. That's the foundation of digital behavioral phenotyping in NHP research.

One caution. Digital phenotyping isn't an automation exercise. The science holds up only if the extracted parameters are biologically relevant, reproducible, and validated for the question at hand. In practice, that means selecting behavioral endpoints before analysis begins, defining reference behaviors, establishing validation procedures, and interpreting AI-derived measurements alongside conventional pharmacology and disease endpoints.

As preclinical research leans harder on quantitative, longitudinal endpoints, behavioral analysis will sit closer to pharmacology, imaging, and biomarker strategies. Markerless 3D tracking, automated behavioral classification, and multimodal translational datasets let you read treatment response at several levels, from drug exposure and biological activity to functional phenotype.

For CNS, pain, pruritus, and other indications where behavior is a core part of the disease phenotype, this gives functional outcomes a more objective measuring stick.

The real opportunity isn't collecting more behavioral data. It's generating better-defined behavioral endpoints that connect to pharmacological mechanisms and translational questions.

This article builds on the Prisys and Bayone webinar, "NHP BehaviorAtlas: High-Fidelity 3D NHP Behavioral Tracking for Drug Development," which covered multi-view 3D tracking, markerless pose estimation, behavioral kinematics, automated behavior classification, and facial expression analysis in NHP research.

For the fuller discussion, both pages are worth a read:

• Original Webinar: NHP BehaviorAtlas Webinar | 3D NHP Behavioral Tracking For Drug Development
• Webinar Recap: Webinar Recap: Advancing Translational NHP Research Through Digital Behavioral Phenotyping

The recap goes deeper on pain, pruritus, neurological disease models, facial expression analysis, and the integration of behavioral endpoints with PK/PD, imaging, pathology, and biomarkers.

For researchers planning NHP efficacy or translational studies: the question isn't whether AI can generate more behavioral data. It's whether that data converts into validated, biologically meaningful endpoints that sharpen how you read treatment response.

Contact Prisys Biotech