Artificial Intelligence Driven Closed Loop Training Systems with Wearable Sensor Integration: A Systematic Scoping Review of Personalized Training Prescription and Training Load Management in Athletes

This pre-registration documents the protocol for a systematic scoping review examining the current state of scientific evidence on artificial intelligence-driven adaptive training systems that integrate wearable sensor data for personalized athletic training prescription and training load management. The protocol was developed in accordance with the PRISMA extension for Scoping Reviews (PRISMA-ScR; Tricco et al., 2018) and is registered prospectively on the Open Science Framework prior to database search execution and data extraction. The convergence of wearable sensor technology and applied machine learning has given rise to a new class of athletic training systems capable of generating structured annual training plans and dynamically adjusting daily prescriptions based on continuous physiological, biomechanical, and subjective wellness data. Devices including GPS trackers, inertial measurement units, heart rate variability monitors, and electromyography sensors now generate high-resolution continuous data streams that, when processed through artificial intelligence and machine learning architectures such as deep learning, reinforcement learning, large language models, and generative adversarial networks, enable closed-loop individualized training decision support at a scale and precision previously unattainable in applied sports science. Classical periodization models, while theoretically well-established, are designed around population averages and predetermined schedules, offering limited capacity to account for day-to-day variability in an individual athlete's physiological state, accumulated fatigue, sleep quality, or subjective readiness. AI-driven systems that integrate wearable data represent a direct methodological response to this limitation. Despite growing commercial adoption and an expanding body of technical literature, the scientific evidence base remains fragmented and methodologically heterogeneous. Comparative trials examining AI-generated versus human-designed training plans are scarce, and where they exist, they vary substantially in study design, participant population, outcome measurement, and intervention duration. Model interpretability poses a persistent challenge, as high-performing deep learning architectures frequently produce accurate predictions without providing the transparent reasoning that athletes and coaches require to build trust in algorithmic recommendations. Critical questions surrounding data standardization, athlete privacy, and the governance of autonomous systems in high-stakes performance contexts remain insufficiently addressed in the literature. These gaps collectively define both the necessity and the scope of the present systematic synthesis. The central purpose of this review is to systematically map and synthesize empirical and technical evidence on AI-driven adaptive training systems that integrate wearable or sensor-derived data, with a particular focus on their capacity to generate annual training plans and adjust daily prescriptions in response to incoming physiological data and athlete-reported wellness feedback. The guiding research question was structured using the CIMO framework (Context, Intervention, Mechanism, Outcome) as established for systematic reviews in management and information systems research by Denyer and Tranfield (2009). The Context dimension encompasses competitive and physically active athletic populations engaged in organized training, spanning recreational to elite competitive levels. The Intervention dimension refers to AI and machine learning systems, including but not limited to deep learning, reinforcement learning, large language models, and generative adversarial networks, that generate structured annual training plans and adjust daily training prescriptions based on wearable device data and athlete-reported wellness feedback. The Mechanism dimension describes adaptive algorithms processing physiological, biomechanical, and subjective wellness data through closed-loop feedback architectures, enabling real-time or near-real-time recalibration of training load. The Outcome dimension concerns evidence of improved training plan quality relative to human-designed plans and effective training load and recovery balance in competitive athletic populations. Stated directly, the central research question asks whether AI-driven systems that generate annual training plans and adjust them daily based on wearable data and user feedback produce training prescriptions of superior quality and more effective load management compared to classical, non-adaptive training methods. This review is designed as a systematic scoping review following PRISMA-ScR as the primary reporting framework. The scoping format was selected over a conventional systematic review due to the fundamental heterogeneity of study designs in this emerging field, which encompasses randomized controlled trials, observational cohort studies, system architecture proposals, algorithm validation papers, and existing systematic reviews. This heterogeneity precludes meaningful quantitative pooling through meta-analysis. The research question, while directional and outcome-focused, is addressed through narrative synthesis and evidence mapping rather than effect-size aggregation, consistent with the purpose of scoping methodology in emerging technology domains as described by Munn et al. (2018). A comprehensive multi-database literature search will be conducted across Web of Science and Scopus as primary databases, supplemented by Google Scholar, SciSpace, and PubMed as secondary sources. The search will be restricted to publications from 2020 onwards, reflecting the period of meaningful convergence between advanced wearable sensor technology and applied machine learning in sports science. The search strategy is organized around three conceptual domains: AI and machine learning technology; training prescription, periodization, and adaptive load management; and athletic populations and performance outcomes. All domain terms will be combined using Boolean operators and adapted to the syntax requirements of each respective database platform. Studies will be included if they satisfy all of the following conditions. The study must employ artificial intelligence or machine learning as a primary component of a training system. The study must address training plan generation, exercise prescription, periodization, or adaptive training load management. The study must incorporate wearable or sensor-derived data, including but not limited to GPS, inertial measurement units, heart rate variability sensors, or electromyography, and or subjective feedback measures such as Rating of Perceived Exertion or subjective wellness ratings as system inputs. The study must report at least one outcome relevant to training plan quality, training load balance, athlete readiness, or athletic performance. The study population must consist of athletes or physically active individuals engaged in organized training. The publication must be a peer-reviewed journal article or full conference paper published from 2020 onwards and available in English full text. Studies focused exclusively on clinical rehabilitation without sports application, wearable hardware development without training system integration, physical education in non-competitive school contexts, or AI applied solely to talent identification or match tactics without any prescriptive output will be excluded. Given the heterogeneity of included study designs, a narrative synthesis approach will be employed, organized around a thematic coding framework. The framework addresses four domains: the AI and machine learning architectures used for training prescription and plan generation; the methods of wearable and sensor data integration within closed-loop feedback systems; the empirical outcomes relating to training plan quality and training load management in competitive athletic populations; and the implementation challenges, interpretability limitations, and research gaps identified across the literature. Where the evidence base permits, a GRADE-informed evidence mapping will characterize the overall strength and certainty of available evidence across the primary outcome domains. This review is expected to produce a comprehensive characterization of the technological architectures underlying AI-driven closed-loop training systems, a synthesis of empirical evidence on the quality of AI-generated training plans relative to human-designed plans and on the effectiveness of wearable-integrated ML systems for training load management, an identification of methodological gaps including the scarcity of high-quality randomized controlled trials and underexplored questions of model interpretability and data governance, and a structured research agenda for future experimental and translational work. By bridging the domains of artificial intelligence, sports physiology, and management decision support systems, this review aims to provide the most comprehensive evidence synthesis to date of closed-loop AI training systems and their practical implications for coaches, sports scientists, system developers, and governing bodies in sport.