AthlePersona: Multimodal Athlete Profile

• AthlePersona is a multimodal, athlete-specific digital twin that integrates kinematics, coaching insights, and biometric data for personalized performance assessment.
• It employs diverse sensing modalities—from wearable sensors to IoT-based vision systems—to capture detailed movement patterns, physiological signals, and psychological cues.
• The framework supports coach-like evaluations through advanced machine learning personalization, real-time retrieval, and a comprehensive dataset of professional athlete profiles.

Searching arXiv for the specified AthlePersona-related papers to ground the article in current literature. AthlePersona denotes a multimodal, athlete-specific representation that encodes movement patterns, technique, physiological condition, psychological state, risk profile, and progression over time. Current arXiv usage suggests two closely related senses. In applied sensing and analytics, AthlePersona functions as a persistent, data-driven digital twin or holistic athlete profile built from wearables, vision, and other telemetry, with the explicit aim of moving beyond repetition counting toward coach-like assessment and individualized monitoring (Ren et al., 2024, Ghosal et al., 26 Jun 2026). In multimodal trait analysis, AthlePersona is also the sports cohort of PersonaX, a dataset of 4181 male professional athletes from seven major sports leagues, combining structured biographical features, one facial image embedding, and large-language-model-inferred Big Five behavior traits (Li et al., 14 Sep 2025).

1. Conceptual Scope and Representational Forms

AthlePersona is organized around the idea that athlete assessment should be person-specific rather than activity-generic. In the coaching-oriented formulation, a profile combines quantitative kinematics, qualitative coaching intelligence, composite protocol-aligned scores, and natural-language feedback. The target is not merely to determine that an athlete is performing a push-up, a run, or a jump, but to encode form degradation, spinal articulation, core rigidity, fatigue, effort, compensation, and consistency in a representation that remains usable across sessions and downstream queries (Ghosal et al., 26 Jun 2026).

A second formulation emphasizes movement analysis as the core of a digital twin. In that framing, AthlePersona is a persistent representation of movement patterns, technique, risk profile, and progression over time, derived from real-time 3D pose estimation, action labels, and biomechanical features. This formulation is especially explicit in track-and-field contexts, where individualized 3D pose histories, phase-specific metrics, and longitudinal risk indicators define the operative persona space (Ren et al., 2024).

The literature also contains compact, interpretable versions of such a representation. "RunnerDNA" defines a five-dimensional movement signature—balance, stride, steering, stability, and amplitude—computed from smartphone multi-sensor data and normalized to $[0,5]$. The purpose is to describe how a person moves, not only what activity is being performed, and to preserve interpretability at the individual level (Yao et al., 2022). An even more compressed representation appears in individual running performance modeling, where best times across distances are described by a rank-3 model

$$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$

with $f_1(s)$ essentially linear in $\log s$, yielding a dominant individual power law $t_a(s)\approx C_a s^{\alpha_a}$ and a three-number summary $(\lambda_1,\lambda_2,\lambda_3)$ that acts as a latent athlete profile (Blythe et al., 2015).

Taken together, these formulations indicate that AthlePersona is not a single fixed architecture. It is better understood as a family of athlete-centered representations ranging from low-dimensional interpretable signatures to multimodal, protocol-aware digital twins. This suggests a common principle: the athlete, not the activity class alone, is the unit of representation.

2. Sensing Modalities and Biomechanical Acquisition

AthlePersona systems are grounded in multimodal sensing. The current literature spans self-powered footwear, IoT pose estimation, smart sportswear, and wearable metabolic monitoring.

Modality	Representative implementation	Primary athlete information
Footwear triboelectric sensing	SG-TENG in the heel region of a running shoe	Gait, cadence, impact intensity, fatigue and tension proxies
IoT pose estimation	IE-PONet with cameras, accelerometers, and gyroscopes	3D pose, action labels, technique and risk indicators
Textile strain sensing	Graphene-based smart compression shirt	Breath-force coordination and left-right muscle activation symmetry
Consumer wearable physiology	Smartwatch and chest-strap modeling pipeline	HR dynamics and instantaneous VO$_2$ trajectories

The SG-TENG line of work provides a self-powered gait interface built from a sodium alginate/gelatin triboelectric layer. The device exhibits an open-circuit voltage of $156.6\ \mathrm{V}$, short-circuit current of $46.9\ \mu\mathrm{A}$, transferred charge of $139.6\ \mathrm{nC}$, and maximum power of $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$0 under optimal load. Its output varies systematically with frequency, force, displacement, and contact area, and when integrated into a running shoe it distinguishes walking, running, and jumping while also supporting capacitor charging through a full-wave bridge (Liang et al., 14 Jul 2025).

Vision-based AthlePersona pipelines use sensor fusion rather than direct force transduction. IE-PONet combines IoT sensing, C3D for spatiotemporal feature extraction, OpenPose for 2D keypoints, and Bayesian optimization for hyperparameter tuning. On NTURGB+D and FineGYM it reports AP$$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$1 scores of $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$2 and $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$3, mAP scores of $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$4 and $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$5, and GFLOPs around $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$6–$$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$7, with ablation studies indicating that temporal features, keypoint localization, and Bayesian tuning each contribute materially to pose-analysis quality (Ren et al., 2024).

Textile sensing targets execution quality at the garment level. The smart sportswear system integrates screen-printed graphene-based strain sensors over the abdomen and bilateral chest, read through an ESP32-based module at $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$8. The quasi-static sensor response reaches a gauge factor of approximately $$\log t(s)=\lambda_1 f_1(s)+\lambda_2 f_2(s)+\lambda_3 f_3(s),$$9, preprocessing uses a $f_1(s)$0 low-pass FIR filter, z-score normalization, and 10 s windows with 50% overlap, and classification is performed by a dual-branch 1D ResNet-18. Across six bench-press execution conditions, the reported overall average accuracy is approximately $f_1(s)$1, with the system distinguishing breathing irregularities and asymmetric muscle exertion (Tang et al., 11 Apr 2025).

Wearable metabolic monitoring extends AthlePersona from biomechanics to physiology. A recent framework models heart-rate dynamics with a physiologically constrained ODE or neural Kalman filter and predicts instantaneous VO$f_1(s)$2 trajectories from consumer-grade wearables. Using over $f_1(s)$3 million HR observations, the best HR model achieves $f_1(s)$4-second interval predictions with mean absolute error as low as $f_1(s)$5 and correlation $f_1(s)$6. The VO$f_1(s)$7 model requires only the initial second of VO$f_1(s)$8 data for calibration and attains mean absolute percentage errors of approximately $f_1(s)$9 across diverse running intensities (Gahtan et al., 30 Apr 2025).

3. Personalization, Adaptation, and Athlete-Specific Inference

A defining property of AthlePersona is personalization. The core technical problem is that inter-subject variability makes one-size-fits-all recognition and assessment unreliable, especially when models are deployed on new athletes.

One strand of work formalizes personalization through subject similarity. In personalized human activity recognition, each subject $\log s$0 is represented by a feature vector $\log s$1 built from physical characteristics, sensor-derived features, or both. Similarity is defined by Euclidean distance and an RBF-style transform,

$\log s$2

This similarity is then used either to weight samples in Personalized Machine Learning or to select subject cohorts in Personalized Deep Learning. The empirical picture is dataset-dependent: on UniMiB-SHAR, Personalized Machine Learning in the hybrid setting reaches $\log s$3 macro average accuracy versus $\log s$4 for the non-personalized deep baseline, whereas on Motion Sense the non-personalized CNN remains stronger (Ferrari et al., 2020).

A second line of work learns athlete- or user-specific embedding spaces. Personalized activity recognition with deep triplet embeddings uses a fully convolutional network and a subject triplet loss

$\log s$5

where anchor, positive, and negative all come from the same subject, but the negative is a different activity. The resulting personalized embeddings outperform both impersonal FCN classification and engineered-feature baselines, reaching $\log s$6 on MHEALTH, $\log s$7 on WISDM, and $\log s$8 on SPAR with the subject triplet variant (Burns et al., 2020).

Meta-learning provides a third personalization regime. Learning-to-learn personalised HAR reframes the task as person-activity classification and constructs each meta-learning task from one person’s support and query sets. Personalised MAML adapts model parameters on the support set of a new person, whereas Personalised Relation Networks condition prediction directly on that support set through learned similarity. Under Leave-One-Person-Out evaluation, these methods outperform conventional deep learning and non-personalized few-shot baselines across exercise, ambulatory, and ADL datasets; practical guidance in the paper places $\log s$9 to $t_a(s)\approx C_a s^{\alpha_a}$0 shots per activity near the favorable accuracy–burden regime (Wijekoon et al., 2020).

Interpretable personalization remains important alongside meta-learning. RunnerDNA shows that a five-indicator profile can still identify activities with accuracy $t_a(s)\approx C_a s^{\alpha_a}$1 and predict the identity of running users better than competing classical models, while the addition of velocity and acceleration improves activity recognition to $t_a(s)\approx C_a s^{\alpha_a}$2. This suggests that interpretable athlete signatures can coexist with, rather than oppose, learned embedding approaches (Yao et al., 2022).

4. Physiological and Psychological State Modeling

AthlePersona is not limited to technique classification. A major theme is the joint inference of physiological and psychological state from movement and wearable signals.

The SG-TENG shoe system explicitly examines gait signatures under neutral, hesitant, and hurried conditions, as well as fatigue progression. Neutral walking is described as having stable amplitude, regular periodicity, and a symmetric pattern. Slowed, uncertain walking shows decreased amplitude, irregular timing, and missing or distorted peaks, while hurried, tense walking exhibits higher frequency and asymmetric signals. Fatigue is similarly described through fluctuating amplitude, irregular cadence, increased noise, and reduced periodicity. The paper does not report statistical validation against heart rate, questionnaires, VO$t_a(s)\approx C_a s^{\alpha_a}$3, or RPE, and therefore these psychophysiological mappings remain qualitative, but the proposed features—mean amplitude, step-interval variability, asymmetry, and abrupt spikes—are directly compatible with state estimators (Liang et al., 14 Jul 2025).

Wearable metabolic modeling provides a quantitatively validated physiological layer. The HR model uses a latent dynamic state, either via

$t_a(s)\approx C_a s^{\alpha_a}$4

in the ODE variant or via a neural Kalman filter with state $t_a(s)\approx C_a s^{\alpha_a}$5. The VO$t_a(s)\approx C_a s^{\alpha_a}$6 model then combines a Kalman-like filtered path, a direct neural regressor, and a blending gate, while using only the initial second of VO$t_a(s)\approx C_a s^{\alpha_a}$7 for initialization. In leave-one-runner-out evaluation on elite runners, the best VO$t_a(s)\approx C_a s^{\alpha_a}$8 model with true HR attains aggregate MAE $t_a(s)\approx C_a s^{\alpha_a}$9, aggregate RMSE $(\lambda_1,\lambda_2,\lambda_3)$0, and aggregate MAPE $(\lambda_1,\lambda_2,\lambda_3)$1. This makes AthlePersona capable of representing not merely external kinematics but cardiovascular and metabolic response dynamics (Gahtan et al., 30 Apr 2025).

Smart sportswear contributes a mesoscopic layer between biomechanics and physiology. Its six classes—Even-Bal, Even-L, Even-R, Uneven-Bal, Uneven-L, and Uneven-R—operationalize breath-force coordination and muscle symmetry during bench press. Grad-CAM visualizations show attention distributed across periodic cycles in coordinated movement and concentrated on irregular or asymmetric segments in faulty execution. This yields a direct representation of respiratory timing, exertion symmetry, and their deviations, which is especially relevant for strength training and rehabilitation personas (Tang et al., 11 Apr 2025).

The combined implication is that AthlePersona can be stratified into at least three interacting subspaces: external biomechanics, internal physiological response, and behaviorally expressed psychological state. Where the literature is quantitative, the persona can support calibration and prediction; where it is qualitative, it presently supports hypothesis generation and coach-facing interpretation.

5. AthlePersona as a Dataset and Causal Analysis Benchmark

In PersonaX, AthlePersona is a released multimodal dataset rather than only a system concept. It contains 4181 male professional athletes drawn from seven major sports leagues worldwide: NBA, NFL, NHL, ATP, PGA, Premier League, and Bundesliga. For each athlete, the released structured features include Id, Height, Weight, Birthyear, Birthmonth, Birthday, League, Latitude, and Longitude; nationality is represented through geocoded latitude and longitude rather than a categorical nationality field. The dataset contains no missing values because rows with missing fields were removed during cleaning (Li et al., 14 Sep 2025).

The visual modality consists of one face embedding per athlete. Raw headshots are collected from official league sources, then embedded using ImageBind into 1024-dimensional representations and transformed through an additional invertible transformation for privacy preservation. The textual modality consists of trait analyses produced by ChatGPT-4o-Latest, Gemini-2.5-Pro, and Llama-4-Maverick from a prompt containing name, gender, league, and country. These texts are embedded using gte-Qwen2-7B-instruct into 3584-dimensional vectors, again with an invertible transformation (Li et al., 14 Sep 2025).

Trait inference is anchored in the Big Five framework, but the paper emphasizes that these are behavior traits inferred from public information rather than clinical personality measures. Each LLM produces scores from $(\lambda_1,\lambda_2,\lambda_3)$2 to $(\lambda_1,\lambda_2,\lambda_3)$3 for Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. Final trait scores are obtained by discarding zeros denoting insufficient information, taking the median of the remaining scores, and rounding fractional medians upward. The authors chose the 3-level numeric scheme after comparing 3- and 5-point scales (Li et al., 14 Sep 2025).

AthlePersona also serves as a benchmark for independence testing and causal representation learning. Across five independence tests—CSQ, GSQ, HSIC, RCIT, and KCI—the strongest dependencies in AthlePersona occur with birth year and league affiliation, while height and weight show more consistent yet moderate associations with behavior traits. In the CRL framework, image embeddings and trait text embeddings are treated as two modalities, and the learned causal graph includes two shared factors $(\lambda_1,\lambda_2,\lambda_3)$4 and $(\lambda_1,\lambda_2,\lambda_3)$5, five image-based latents, and five trait-based latents. The paper interprets several of these latents as “mindset,” “culture,” “confidence,” “emotional stability,” “self-awareness,” “skin tone,” “grooming,” and “facial expressions,” though these labels remain interpretive rather than directly observed variables (Li et al., 14 Sep 2025).

This dataset-centered usage widens the meaning of AthlePersona. It turns athlete profiling into a problem of multimodal representation learning, statistical dependence analysis, and causal discovery under privacy constraints, rather than only biomechanics or coaching automation.

6. System Architectures, Retrieval Layers, and Open Problems

The most fully articulated end-to-end AthlePersona architecture is the agentic framework for holistic athlete profiling. Orchestrated with LangGraph, it uses a Guardrail or Validator, a VLM qualitative biomechanics agent, a CV kinematic tracker, an Assessor or Aggregator, an LLM-as-a-Judge self-correction loop, relational persistence, vector persistence, a Retriever, and an Analyst. The system introduces a $(\lambda_1,\lambda_2,\lambda_3)$6 Smart Grid temporal chunking strategy that reduces VLM calls by over $(\lambda_1,\lambda_2,\lambda_3)$7, and the Judge agent routes low-confidence cases back for re-evaluation or manual review. In a user study with 40 coaches or instructors, mean assessment time per athlete falls from $(\lambda_1,\lambda_2,\lambda_3)$8 min to $(\lambda_1,\lambda_2,\lambda_3)$9 min, agreement with expert SAI rubric rises from $_2$0 to $_2$1, feedback actionability rises from $_2$2 to $_2$3, and SUS usability rises from $_2$4 to $_2$5 (Ghosal et al., 26 Jun 2026).

This agentic design is notable because it adds persistence and retrieval to the AthlePersona stack. Structured metrics are written to Google Sheets or an equivalent relational store, while free-text assessments are embedded and stored in ChromaDB. Natural-language queries such as identifying athletes with high endurance but poor core rigidity are answered through retrieval-augmented generation. AthlePersona therefore functions not only as a sensing representation but also as a searchable knowledge substrate (Ghosal et al., 26 Jun 2026).

Several papers converge on similar deployment requirements. SG-TENG work explicitly proposes wireless transmission, data-driven analytics, multi-sensor fusion, and early warning for fatigue or stress, while also noting absent long-term durability testing, environmental sensitivity, incomplete calibration to absolute force, lack of a wireless module, and qualitative rather than statistically validated psychophysiological mappings (Liang et al., 14 Jul 2025). IE-PONet points toward multimodal data integration and real-time feedback mechanisms, emphasizing edge–cloud collaboration and adaptive hyperparameter optimization for real environments (Ren et al., 2024). Smart sportswear identifies a need for broader participant coverage, smaller and more flexible electronics, broader exercise coverage, and integration with smartphone apps, cloud analytics, LLM agents, and human body digital twins (Tang et al., 11 Apr 2025).

The PersonaX dataset introduces a different class of open problems: demographic restriction to male athletes, concentration in North American and European elite leagues, higher missing or insufficient-information rates for athletes than celebrities, and the ethical need to prohibit high-stakes or commercial deployment. The dataset is framed for non-commercial research, and its obfuscated embeddings and usage guidelines reflect a privacy-preserving orientation (Li et al., 14 Sep 2025).

These converging limitations indicate that AthlePersona remains a research program rather than a closed standard. The literature supports high-fidelity sensing, rapid personalization, natural-language retrieval, and compact latent profiling, but robust deployment still depends on durability, calibration, demographic breadth, multimodal validation, privacy governance, and careful treatment of psychological inference.