Personalized rehabilitation programs are tailored interventions designed to address the unique recovery needs of patients undergoing urologic oncology treatments. These programs integrate various therapeutic modalities, including physical therapy, nutritional guidance, psychological support, and exercise regimens, to enhance postoperative recovery and overall quality of life Nawara and Kashef (19). The customization of these programs ensures that individual patient characteristics, such as age, comorbidities, and baseline functional status, are considered, leading to more effective rehabilitation outcomes Feng et al. (20). Recent studies have highlighted the efficacy of prehabilitation—interventions initiated before surgery—in improving surgical outcomes for urologic cancer patients. A systematic review published in European Urology examined the impact of prehabilitation exercise programs on presurgical cardiopulmonary fitness measures Khan et al. (21). The findings indicated that such programs effectively enhance presurgical fitness, potentially leading to better postoperative recovery. However, the review also emphasized the need for improvements in program design and reporting to conclusively determine their impact on surgical outcomes. Incorporating technology into personalized rehabilitation has gained momentum Rocco et al. (22). Mobile applications and wearable devices offer platforms for delivering tailored exercise programs, monitoring patient progress, and facilitating real-time feedback. A study discussed in European Urology Focus explored the emerging role of these technologies in prehabilitation for urologic oncology patients Cabrera-Sánchez et al. (23). The study found that wearable devices could improve access to prehabilitation programs, reduce the need for in-person visits, and allow for continuous monitoring, thereby enhancing patient engagement and adherence. Moreover, the integration of personalized rehabilitation programs within Enhanced Recovery After Surgery (ERAS) protocols has been associated with reduced morbidity and improved recovery times. A study in the European Journal of Surgical Oncology evaluated the application of ERAS protocols in patients undergoing radical cystectomy Fu et al. (24). The study demonstrated that implementing ERAS protocols, which include personalized rehabilitation components, significantly reduced the length of hospital stay and postoperative complications, underscoring the value of such comprehensive approaches in urologic oncology care.

Robotic-assisted surgical techniques have revolutionized the field of urologic oncology by offering minimally invasive options for procedures such as prostatectomy, nephrectomy, and cystectomy. These techniques utilize robotic systems to enhance surgical precision, reduce operative times, and minimize patient morbidity Argyriou et al. (25). The adoption of robotic-assisted surgery has been associated with improved oncological outcomes and faster recovery periods compared to traditional open surgeries Nawara and Kashef (26). A notable advancement in this domain is the development of robot-assisted retroperitoneal lymph node dissection (RPLND) for metastatic testicular cancer. Traditional open RPLND is associated with significant morbidity and extended recovery times Lee et al. (27). However, as discussed in a Cleveland Clinic podcast, the robotic-assisted approach offers excellent cancer control with vastly improved recovery, making it a preferred option among urologic oncologists for select patients. The benefits of robotic-assisted surgery extend beyond oncological outcomes Yadalam et al. (28). A study highlighted by University College London demonstrated that robot-assisted surgery for bladder cancer removal enables patients to recover more quickly and spend significantly less time in the hospital compared to traditional open surgery. This minimally invasive approach not only enhances patient recovery but also reduces healthcare costs associated with prolonged hospital stays Hsia et al. (29). Furthermore, the continuous evolution of robotic technology has led to the development of single-port robotic systems, allowing surgeons to perform complex urological procedures through a single incision. This advancement minimizes surgical trauma and enhances cosmetic outcomes, further solidifying the role of robotic-assisted techniques in modern urologic oncology.

Recent clinical studies underscore the growing adoption of robotic-assisted systems in various urologic procedures, reflecting a transformative shift in surgical practices. For example, advancements in robot assisted radical prostatectomy and nephrectomy have been linked with improved oncological control, reduced blood loss, and faster patient recovery times Di Bello et al. (30) Carilli et al. (31). Moreover, the development of next-generation robotic platforms, including single-port systems and AI-enhanced visualization, has broadened surgical precision and accessibility Di Bello et al. (32) Falkenbach et al. (33). These trends indicate not only increasing trust in robotic solutions among urologic surgeons but also the urgent need to develop parallel innovations in post-surgical rehabilitation to match the precision and individualization offered intraoperatively. Our proposed framework addresses this gap by leveraging AI-driven rehabilitation aligned with the evolving robotic surgical landscape.

The integration of technology into rehabilitation programs has opened new avenues for personalized patient care in urologic oncology recovery Yeh and Kashef (34). The use of mobile applications, wearable devices, and adaptive rehabilitation systems facilitates continuous monitoring, personalized exercise regimens, and real-time feedback, thereby enhancing patient engagement and adherence to rehabilitation protocols Zhang et al. (35). A study published in the International Journal of Human-Computer Interaction introduced a personalized sports health recommendation system assisted by a Q-Learning algorithm. This system leverages artificial intelligence to provide tailored exercise recommendations based on individual user data, promoting effective and safe rehabilitation practices Ko et al. (36). The adaptive nature of the system ensures that exercise intensity and type are aligned with the patient’s current health status and recovery progress Forouzandeh et al. (37). In the realm of prehabilitation, mobile app-based programs have shown promise in preparing patients for surgery and aiding in postoperative recovery. A prospective nonrandomized study discussed on UroToday evaluated the impact of a personalized mobile app designed for patients undergoing radical prostatectomy Collà Ruvolo et al. (38). The findings suggested that the app-based program improved patient outcomes by facilitating better preparation and recovery processes, highlighting the potential of digital health interventions in surgical care. The development of adaptive rehabilitation systems, such as the one by Kaunas University of Technology (KTU), offers personalized recovery experiences Ruvolo et al. (39). These systems adjust the rehabilitation exercises based on the user’s capabilities, ensuring a safe training process and avoiding overloading. Such technology driven approaches cater to the individual needs of patients, making rehabilitation more effective and accessible. The integration of technology in rehabilitation not only personalizes the recovery process but also addresses logistical challenges by reducing the need for in-person visits. This is particularly beneficial for patients with mobility issues or those residing in remote areas, ensuring that they receive consistent and effective rehabilitation support throughout their recovery journey Ruvolo et al. (40).

Emerging biomedical evidence also highlights the systemic benefits of exercise, extending beyond musculoskeletal recovery to include metabolic regulation. Notably, recent research suggests that structured physical activity can mitigate lipidic imbalance—a common issue in postoperative oncology patients that exacerbates cardiovascular risk and systemic inflammation Di Bello et al. (41). By integrating exercise interventions tailored to each patient’s physiological profile, our proposed rehabilitation system could contribute to the normalization of lipid profiles and promote comprehensive recovery. This additional metabolic perspective reinforces the multi-dimensional value of personalized exercise prescription in urologic oncology care.

Sports analytics has emerged as a crucial interdisciplinary field that integrates data science, machine learning, and domain knowledge to enhance decision-making processes in sports. This section provides an overview of the methodologies and principles that underpin our approach to sports analytics. A core challenge in sports analytics lies in the complex and dynamic nature of sports data, which includes structured tabular records and unstructured data. Traditional methods rely on statistical models, whereas modern techniques leverage deep learning and reinforcement learning to capture intricate patterns and make predictive or prescriptive decisions. Our approach aims to bridge these methodologies, incorporating domain-specific constraints and real-time adaptability.

In Section 3.2, we formalize the sports analytics problem by defining a mathematical representation of key performance metrics, player behaviors, and game dynamics. This enables a structured formulation that facilitates computational modeling. In Section 3.3, we introduce our novel model, which integrates feature engineering, neural networks, and optimization techniques to enhance predictive accuracy and interpretability. Section 3.4 presents our strategic framework that refines decision-making by incorporating real-time adjustments and uncertainty quantification, ensuring robust and adaptive analytics solutions. By systematically integrating data-driven models with domain expertise, our framework seeks to advance sports analytics beyond conventional statistical approaches, paving the way for more effective performance evaluation, game strategy optimization, and injury prevention techniques.

In this section, we formalize the problem of sports analytics by defining a mathematical framework that encapsulates key aspects such as player performance modeling, game dynamics, and decision-making processes. Let T denote the total duration of a game, discretized into time steps t ∈ { 1 , 2 , … , T } . Let   P = { p 1 , p 2 , … , p N } be the set of players, where each player p i is associated with a feature vector x i ( t ) ∈ ℝ d at time t , representing attributes such as velocity, stamina, and skill level.

We define the game state at time t as (Equation 1):

where X ( t ) = [ x 1 ( t ) , … , x N ( t ) ] ∈ ℝ N × d is the player state matrix, Y ( t ) ∈ ℝ N × m represents observed actions, and G t ∈ ℝ k encodes global game context such as score, possession, and remaining time.

The transition of player states over time can be modeled as (Equation 2):

where u i ( t ) is the control input, and η i ( t ) is a stochastic noise term representing uncertainties in movement.

Each player selects an action a i ( t ) ∈ A based on a policy function (Equation 3):

where Q ( S t , a ) is the estimated value of taking action a in state S t . This follows the standard softmax decision model.

The probability of winning W given the game trajectory τ = { S 1 , S 2 , … , S T } is modeled as (Equation 4):

where σ ( · ) is the sigmoid function, w t are learnable weights, and ϕ ( S t ) is a feature transformation function capturing game impact.

For each player, we define an individual performance score R i based on contributions to key events (Equation 5):

where ℰ represents event types, α e are predefined importance weights, and I e , i ( t ) is an indicator function for event occurrence.

The overall team performance is given by (Equation 6):

where C ( t ) represents team fatigue cost, and λ is a regularization coefficient balancing performance and stamina.

Given historical game data D = { S t , A t , R t } t = 1 T , our objective is to learn a predictive function F such that (Equation 7):

where θ are model parameters optimized via (Equation 8):

This formulation provides a comprehensive mathematical foundation for sports analytics, capturing player behaviors, game evolution, and performance evaluation. The subsequent sections will introduce our novel model and strategy to enhance prediction accuracy and decision-making effectiveness.

In this section, we introduce the Dynamic Sports Performance Network (DSPN), a novel architecture designed to analyze and predict player performance and game outcomes in the context of sports analytics(As shown in Figure 1 ). DSPN incorporates temporal player dynamics, strategic player interaction, and contextualized game understanding into a unified modeling paradigm.

To enhance the predictive depth and strategic awareness of the Dynamic Sports Performance Network (DSPN), we propose the Adaptive Game Intelligence Strategy (AGIS), a novel integration of decision theory, reinforcement learning, and uncertainty modeling(As shown in Figure 3 ). AGIS introduces three key innovations that enable real-time tactical adjustment, opponent-aware adaptation, and risk-sensitive decision-making.

The MovieLens Dataset González et al. (42) is a widely used benchmark for recommendation systems, particularly in movie rating and recommendation tasks. It contains various versions, ranging from 100,000 ratings (MovieLens 100K) to over 25 million ratings (MovieLens 25M), contributed by thousands of users on thousands of movies. Each dataset includes metadata such as movie titles, genres, and timestamps of user ratings. The diversity in user preferences, temporal behavior, and genre distribution makes it ideal for evaluating collaborative filtering and matrix factorization techniques. The dataset is preprocessed and anonymized to facilitate reproducible research in recommender systems and machine learning. The Yelp Dataset Kumar et al. (43) is a large-scale dataset curated from Yelp’s business platform, used extensively for research in natural language processing, sentiment analysis, and recommendation systems. It includes millions of reviews, business information, user profiles, and ratings across various domains like restaurants, shops, and services. Each review contains textual content, star ratings, and timestamps, providing a rich context for analyzing user sentiment and behavior. The dataset also includes social network information among users, allowing for studies involving social influence in recommendations. The Yelp Dataset presents realistic challenges such as noisy text, user bias, and data sparsity. The Jester Dataset Senyurek and Kevric (44) is a benchmark dataset for real-valued collaborative filtering, focusing on joke recommendation. It consists of ratings from users on a fixed set of jokes, with continuous ratings ranging from -10 to +10, rather than discrete integers. The dataset includes over 1.7 million ratings from thousands of users, making it one of the few publicly available datasets for evaluating real-valued recommender systems. Unlike sparse datasets, Jester features a dense user-item matrix, which allows detailed analysis of preference modeling and rating prediction. Its unique characteristics make it suitable for evaluating both traditional and deep learning-based recommendation algorithms. The MIMIC-III Dataset Budrionis et al. (45) is a large, publicly available dataset for medical research, containing de-identified health records of over 40,000 critical care patients from the Beth Israel Deaconess Medical Center. The dataset includes structured data such as demographics, vital signs, laboratory tests, medications, and ICD codes, as well as unstructured clinical notes. MIMIC-III supports a wide range of tasks including patient outcome prediction, disease classification, and clinical decision support. Its comprehensive and real-world nature makes it a cornerstone dataset for machine learning research in healthcare, particularly in the context of electronic health records (EHRs).

In our experiments, we adopt PyTorch as the primary deep learning framework due to its flexibility and extensive support for GPU acceleration. All models are trained using NVIDIA A100 GPUs with 40GB memory, enabling efficient handling of spatiotemporal data and large-scale batch processing. The input video frames are uniformly resized to 224 × 224 pixels and normalized channel-wise using the standard ImageNet statistics (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]). To capture diverse temporal patterns, we apply a segment-based temporal sampling strategy, where each video is evenly divided into T segments and a representative frame or short snippet (depending on model type) is sampled from each segment. This ensures both temporal coverage and computational tractability. The number of frames per input video is set to either 8 or 16, balancing recognition performance and GPU memory consumption. For spatial data augmentation during training, we use a combination of random horizontal flipping, random resized cropping (with scale range from 0.8 to 1.0), and color jittering (adjusting brightness, contrast, saturation, and hue), which helps improve generalization and robustness to environmental variations. Training is performed using stochastic gradient descent (SGD) with a momentum coefficient of 0.9 and weight decay set to 5×10⁻⁴. The learning rate is initialized at 0.01 and decayed by a factor of 10 at epochs 30 and 60, following a step decay schedule. Training concludes at epoch 100. We use a fixed mini-batch size of 64 across all datasets and settings for consistency. Dropout with a rate of 0.5 is applied before the final fully connected layer to mitigate overfitting. During evaluation, we report both top-1 and top-5 classification accuracy metrics. For inference, each test video is processed using center cropping and 10 uniformly sampled clips, whose prediction scores are averaged to obtain the final decision. For models that incorporate 3D spatiotemporal convolutions, we initialize the 3D layers with pretrained weights from Kinetics-400 to accelerate convergence and improve accuracy. For transformer-based models, we either train from scratch or fine-tune from ImageNet-pretrained 2D backbones, depending on the availability of pretraining resources. All baseline models are re-implemented and trained under the exact same data pipeline and optimization schedule to ensure fair and reproducible comparisons. We also measure inference throughput in terms of frames per second (FPS) on a single A100 GPU, reporting average FPS over 100 test videos to assess runtime efficiency under deployment conditions ( Algorithm 1 ; Equations 35–42).

Tables 1 , 2 present the comparison of our method with several state-of-the-art (SOTA) models on UCF-101, HMDB-51, ActivityNet, and Kinetics datasets. The comparison is conducted using precision, recall, F1-score, and NDCG metrics, ensuring a comprehensive evaluation of model performance.

Our method consistently outperforms the baseline models across all datasets. On the UCF-101 dataset, our model achieves an F1-score of 87.83, surpassing BERT4Rec (84.65) and SASRec (83.91), demonstrating the effectiveness of our approach in recognizing complex actions. The high NDCG score of 90.34 further indicates the superior ranking quality of our predictions. Similarly, on HMDB-51, our method obtains an F1-score of 81.85, outperforming BERT4Rec (78.66) and SASRec (77.78), highlighting its robustness in more challenging action categories. The superior recall score (86.45 on UCF-101 and 81.10 on HMDB-51) suggests that our approach captures relevant action instances with higher accuracy. For the ActivityNet dataset, our method achieves an F1-score of 86.15, which is significantly higher than BERT4Rec (83.52) and SASRec (82.42). The NDCG score of 88.93 reflects the improved ranking performance, indicating the model’s ability to prioritize correct action predictions effectively. On Kinetics, our approach reaches an F1score of 80.52, again outperforming BERT4Rec (78.05) and SASRec (77.23). The higher precision (81.10) compared to the best baseline (78.58 from BERT4Rec) suggests that our model minimizes false positive action predictions more effectively. The consistent improvement across multiple datasets can be attributed to several factors. Our method employs an enhanced feature representation strategy, allowing for better differentiation between visually similar actions. Our temporal modeling captures long-range dependencies, which is particularly beneficial for datasets like ActivityNet and Kinetics, where actions span longer durations. Our model benefits from an optimized training pipeline with advanced regularization techniques, reducing overfitting and improving generalization. The use of pretraining on large-scale datasets, followed by fine-tuning, enables our model to leverage transferable knowledge, resulting in superior performance across different benchmarks.

In Figures 5 , 6 , our method achieves the highest scores across all datasets, outperforming existing models in action recognition tasks. The results validate the effectiveness of our approach in capturing complex action patterns, leading to state-of-the-art performance in precision, recall, F1-score, and NDCG. These findings highlight the strength of our proposed model in both short-term and long-term action understanding.

To further analyze the contribution of each component in our proposed model, we conduct an ablation study on the UCF-101, HMDB-51, ActivityNet, and Kinetics datasets. The results are presented in Tables 3 , 4 . We evaluate three main ablations: removing Spatio-Temporal Player Encoding, Multi-Agent Tactical Learning, and Opponent Behavior Anticipation. The full model (Ours) consistently achieves the best performance, demonstrating the importance of each component.

From the results on the UCF-101 dataset, removing Spatio-Temporal Player Encoding leads to a drop in F1-score from 87.83 to 83.90, indicating that Spatio-Temporal Player Encoding plays a crucial role in improving classification accuracy. Similarly, without Multi-Agent Tactical Learning, the F1-score decreases to 85.03, while removing Opponent Behavior Anticipation results in a more significant drop to 82.85. The NDCG score follows a similar trend, confirming that these components contribute significantly to ranking quality. The performance drop is even more pronounced on HMDB-51, where excluding component Spatio-Temporal Player Encoding, Multi-Agent Tactical Learning, or Opponent Behavior Anticipation leads to a decrease in F1-score from 81.85 to 78.80, 79.82, and 77.98, respectively, highlighting the importance of these features in recognizing complex human actions. On the ActivityNet dataset, our full model achieves an F1-score of 86.15, while removing Spatio-Temporal Player Encoding, Multi-Agent Tactical Learning, or Opponent Behavior Anticipation results in performance drops to 82.94, 84.21, and 82.34, respectively. The NDCG score also decreases significantly, suggesting that each component plays a crucial role in effective action recognition. Similarly, on the Kinetics dataset, the full model achieves an F1-score of 80.52, but removing Spatio-Temporal Player Encoding, Multi-Agent Tactical Learning, or Opponent Behavior Anticipation leads to decreases in performance, with the largest drop observed when component Opponent Behavior Anticipation is removed (76.91 F1-score). These results indicate that each component of our model contributes to capturing important action cues and improving recognition accuracy.

The effectiveness of our method can be attributed to several factors. Spatio-Temporal Player Encoding enhances feature representations, allowing the model to distinguish between visually similar actions more effectively. Component Multi-Agent Tactical Learning improves temporal modeling, capturing long-range dependencies critical for complex activities. Component Opponent Behavior Anticipation refines the final predictions, ensuring that the model maintains high precision and recall. The ablation results validate our architectural choices and demonstrate that the combination of these components leads to a more robust and accurate action recognition system. In Figures 7 , 8 , our ablation study confirms that each module in our proposed framework plays a vital role in achieving state-of-the-art performance. By systematically removing each component, we observe significant drops in key evaluation metrics, reaffirming the necessity of our design choices in improving action recognition across different datasets.