Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

666 https://internationalpubls.com 

Design of an Improved Model for Multimodal Data Fusion Using 

XGBoost-LSTM-CNN and Proximal Policy Optimizations 

 

Vijaya Kamble 1*   Dr. Sanjay Bhargava 2 

1 Ph.D Research Scholar, Department of Computer Science and Engineering, Mansarovar Global University, Billkisganj, 

Sehore, Madhya Pradesh-466001. vijaya.kamble@ghrua.edu.in, 

2 Research Guide, Department of Computer Science, Faculty of Engineering and Technology, Mansarovar Global 

University. sanjaybhargava78@gmail.com 

 

Article History: 

Received: 15-07-2024 

Revised: 05-09-2024 

Accepted: 28-09-2024 

Abstract:  

Applications in healthcare, finance, and real-time sensor applications call for much stronger 

demands of improving accuracy and efficiency in the analysis of multimodal data. Current 

methods have deficiencies in fusing multiple types of data such as time-series, spatial, and 

categorical data, arising due to limitations in capturing sequential dependencies, spatial 

patterns, and feature importance simultaneously. It also discusses these challenges through a 

novel solution based on a hybrid ML-DL approach, with an integration of reinforcement 

learning and advanced probabilistic models. First, the method that comes to mind is the 

XGBoost-LSTM-CNN Hybrid Model, wherein three drivers of performance improvements 

come together, namely, XGBoost, with much-appreciated capability for outstanding handling 

of structured tabular data, LSTM due to its proficiency in capturing long-term temporal 

dependencies, and the strength of CNN in spatial feature extraction. These results improve 

the predictive accuracy for multimodal datasets. After that, further enhanced multimodal 

fusion by the CAMT selectively pays attention to the critical features across the modalities, 

enhancing the accuracy of the contextual time-series predictions. Subsequently, PPO enables 

real-time model adaptation by dynamic optimization of model parameters and improvement 

of predictions through continuous learning. KF-BNN reduces noise and uncertainty in time-

series data by fusing the filtering capability of a Kalman filter with probabilistic modeling 

via Bayesian neural networks to give reliable predictions. It also provides federated learning 

via FedAvg with online gradient descent for distributed model training in a manner that 

ensures privacy, continuous model updates without having to centralize sensitive data 

samples. These approaches show significant improvements along many axes, with accuracy 

improvements of up to 4.5% and MAE reductions of 12% relative to the baseline models. 

Proposed models provide a valid framework for analyzing multimodal data, thereby 

increasing precision, recall, and overall adaptiveness in dynamic real time, hence pushing the 

state-of-the-art in multimodal fusion and timeseries prediction. This research probably will 

influence those areas dependent on distributed, multimodal, and sequential data samples. 

Keywords: Multimodal Fusion, XGBoost, LSTM, Proximal Policy Optimization, Time-

Series Predictions 

 

1. Introduction 

Modern streams of data have become so complex and diverse, from time-series to audio, video, and 

sensor data; thus, they inherently demand advanced models that can handle sets of information with 

multiple modes. Traditional machine learning and deep learning methods are effective in individual 

mailto:sanjaybhargava78@gmail.com


Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

667 https://internationalpubls.com 

domains but usually fail in the integration and joint analysis of such heterogeneous data types. The 

current architectures are not able to capture underlying temporal dependencies, spatial correlations, 

and varied structures of real-world applications in health care, finance, or autonomous systems. Very 

often, approaches for multimodal data analytics employ models tailored for either tabular, sequential, 

or spatial data samples. For instance, models like LSTM [4, 5, 6] are quite effective for time-series 

analysis owing to the presence of long-term dependencies, while CNNs act as strong tools to extract 

spatial features from image sequences or sensor data samples. However, most of these methods, when 

applied separately, can hardly ensure robust performance for a fully integrated multimodal approach 

in dynamic and noisy environments. Besides, classic machine learning algorithms, such as decision 

trees and gradient boosting, of which examples are models like XGBoost, are very good at structured 

tabular data and able to show feature importance, while rather limited in sequential data modeling. 

Hybrid approaches can combine the strengths of various models and hence have a great potential to 

resolve these limitations. 

It introduces an overall model that integrates XGBoost with the deep learning mechanisms, LSTM and 

CNN, to improve multiple modal fusion. This paper then goes on to propose Proximal Policy 

Optimization, reinforcement learning that dynamically adjusts model parameters in real-time to adapt 

to continuous changes in data. Complementing these methods is the adoption of KFBNN - Kalman 

Filter-based Bayesian Neural Network, improving the quality of the predictions by filtering noise and 

including the estimation of uncertainty. Finally, federated learning using FedAvg provides model 

adaptations across different edge devices in a distributed fashion, enabling data privacy and continuous 

learning from decentralized sources. In this way, the proposed hybrid framework significantly 

improves the shortcomings of traditional techniques of multimodal fusion and yields improved 

precision, recall, and overall prediction accuracy. The model proposed will provide a robust solution 

for real-time multimodal data analytics, which relates to various applications such as health monitoring 

systems, financial forecasting, and smart cities; in all of these, multiple data sources have to be 

integrated seamlessly to result in an optimum decision-making process. 

2. In-depth review of existing models for Time Series Analysis 

Time series forecasting has been lately the focus of a wide variety of sciences, while most of the latest 

contributions are focused on improving the accuracy of the predictions, deal with multivariate data 

sets, and integrate multiple modal data. This review will look at some recent work in the domain and 

place the proposed hybrid model, namely XGBoost-LSTM-CNN, into the perspective of larger-scale 

multimodal and time-series prediction studies. A fault prediction model for electromagnetic launch 

systems is proposed by Junyong et al. using time-series analysis and neural networks. While this model 

works well in structured data, it has applications only in specific domains where time-series data forms 

the single focus, for example, railguns. In addition, Kim and Kim  have suggested the use of 

convolutional transformer models for multivariate time series prediction as a way to enhance feature 

extraction by incorporating transformer mechanisms. While transformers are powerful in the tasks of 

sequence-to-sequence modeling, their computational complexity also renders them less feasible than 

the reinforcement learning part of the proposed model for real-time applications. In, Zeng et al. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

668 https://internationalpubls.com 

proposed a fuzzy time series prediction approach to fault prediction of large scale pulse capacitors. 

While fuzzy logic indeed adds robustness in uncertain environments, this model incorporates BNN for 

uncertainty estimation, thus making a stronger filtration for noise and hence better probabilistic 

reasoning possible. Feng et al. [4] contributed an MTL framework that captures both dynamic-shared 

and specific patterns for chaotic timeseries prediction. The approach given by them is efficient in 

multitasking environments; however, the hybrid model proposed here is more versatile as it uses CNN 

for spatial data and LSTM for sequential pattern in case of multimodal data fusion. Shen et al. [5] 

presented multivariate time-series forecasting using elastic net and high-order fuzzy cognitive maps 

that are used to predict EEG signals. While the 1D-CNN used in their work improves the process of 

spatial-temporal feature extraction, integration of CNNs within the proposed model for handling 

spatial data with XGBoost provides a wider solution. Zhou et al. [6] addressed industrial process 

prediction under limited data using transfer learning. Their approach is for a few-shot setting and thus 

is complementary to the transfer learning-based extension that can be made to the proposed model 

toward real-time adaptability. 

Yin et al. [7] developed a GAN with multiple attention for time-series prediction. While GANs are 

truly effective in the generation of synthetic data, their instability in training is in contrast to stability, 

which the use of PPO introduces in the model proposed here. Yi et al. [8] proposed an intergroup 

cascade broad learning system with optimized parameters for chaotic time-series prediction; hence, 

the emphasis was more on the optimization of parameters. While this could be powerful for chaos-

related tasks, the broader adaptability of PPO in the proposed model provides superior performance 

across multimodal domains. Chen and Sun [9] introduced Bayesian temporal factorization for 

multidimensional timeseries prediction. Their use of probabilistic models aligns closely with the 

Kalman filter-enhanced BNN in the proposed model. However, the latter is more suitable for real-time 

prediction tasks due to fusion with XGBoost, LSTM, and CNN. There is a VAM simulator developed 

by Mubang and Hall  used for social media time-series prediction. Mubang and Hall focus on social 

networking in their work, but the core of their work shows the importance of network-based analysis 

toward time series. The model could be useful at different areas of social media, health care, and 

finance. In such a context, Ma et al.  proposed the method for short-term traffic flow prediction using 

LSTM and BiLSTM on an urban traffic system. Their method is restricted only to the traffic data, 

whereas this proposed model handles multimodal data because it integrates CNNs for capturing the 

spatial tasks and LSTMs for temporal tasks. Hence, this network will be more robust in real-world 

scenarios. Wang et al.  [12] employed an information granules-based BP neural network aimed at long-

term time series prediction. While effective for granular data, their model does not have the ability of 

multimodal fusion like the proposed architecture, which integrates the various data types to give an 

encompassing prediction. Maaroofi et al.  proposed a time-series prediction using the ensemble data 

autocorrelation forecasting method. Though ensemble methods improve the prediction accuracy, the 

integration of PPO enables real-time adaptability of the proposed hybrid model and improves model 

efficiency in dynamic environments. Ren et al. [14] addressed a multivariate utility time-series 

representation and prediction with a focus on the modeling of sensory data. The proposed model here 

provides an extension to include federated learning in the case of distributed datasets-a vital factor for 

preserving privacy in utility demand predictions within smart cities. Yuan et al. [15] presented a joint 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

669 https://internationalpubls.com 

spatiotemporal feature learning framework for multivariate time-series prediction using fuzzy 

cognitive maps and sparse autoencoders. Their approach is somewhat related to the proposed model 

through the use of CNNs to extract spatiotemporal features, although this model furthers their approach 

with the addition of XGBoost for tabular data, lending it a more holistic approach in multimodal fusion. 

On the whole, though contributive in large measures within certain specific domains, those pertaining 

to time-series and multimodal data predictions, the integration of XGBoost, LSTM, CNN, PPO, and 

Bayesian Neural Networks has made the proposed model far more comprehensive, scalable, and 

adaptable. It not only improves the accuracy of prediction and real-time adaptability but extends to 

construct wide applicability on a range of multimodal datasets: from healthcare to finance, sensor 

networks, and smart cities. 

3. Proposed Design of an Improved Model for Multimodal Data Fusion Using XGBoost-LSTM-

CNN and Proximal Policy Optimizations 

This section provides a comprehensive solution for multimodal data fusion, time-series prediction, and 

dynamic adaptation by leveraging the strengths of XGBoost, LSTM, CNN, and reinforcement learning 

via PPO. Each constituent addresses a specific part of the general complex problem of multimodal 

data: XGBoost for structured tabular data handling in an efficient manner, LSTM for capturing long-

term dependencies in sequential data, CNN for extracting spatial features, and PPO for dynamic 

adaptability in runtime scenarios. This model is designed to enhance the prediction accuracy and 

robustness by leveraging synergies among the said techniques. It also provides an advanced Bayesian 

technique to enable online learning continuously and filtering noises. XGBoost will perform the initial 

process on static or non-sequential tabular features. In this approach, a loss function L(θ) is defined as 

a regularized objective balancing model accuracy and model complexity. The regularized loss function 

is given via equation 1, 

𝐿(𝜃) =∑𝑙(𝑦𝑖, 𝑦′𝑖(𝜃)) + 𝛺(𝜃)… (1)

𝑛

𝑖=1

 

Where, 𝑙(𝑦𝑖, 𝑦′𝑖(𝜃)) represents the loss between true values yi and predicted values y'i(θ), while Ω(θ) 

represents the regularization term, which helps in preventing overfitting by penalizing model 

complexity levels. The output of XGBoost gives feature importance and a baseline prediction, which 

is combined further with the deep learning components. Among the considered models, the LSTM 

network had the most important role to handle temporal dependencies in sequential data. The LSTM 

will take as input sequences Xt, either time series or sensor data, and will keep track of a memory state 

Ct across temporal sets of instances. The cell and hidden state update can be done by the following 

equations 2, 3, 4, 5 & 6, 

𝑓𝑡 = 𝜎(𝑊𝑓 ⋅ [ℎ(𝑡 − 1), 𝑋𝑡] + 𝑏𝑓)… (2) 

𝑖𝑡 = 𝜎(𝑊𝑖 ⋅ [ℎ(𝑡 − 1), 𝑋𝑡] + 𝑏𝑖)… (3) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

670 https://internationalpubls.com 

𝐶𝑡 = 𝑓𝑡 ⊙ 𝐶(𝑡 − 1) + 𝑖𝑡 ⊙ 𝑡𝑎𝑛 ℎ(𝑊𝑐 ⋅ [ℎ(𝑡 − 1), 𝑋𝑡] + 𝑏𝑐)… (4) 

𝑜𝑡 = 𝜎(𝑊𝑜 ⋅ [ℎ(𝑡 − 1), 𝑋𝑡] + 𝑏𝑜)… (5) 

ℎ𝑡 = 𝑜𝑡 ⊙ 𝑡𝑎𝑛 ℎ(𝐶𝑡)… (6) 

 

Figure 1. Model Architecture of the Proposed Analysis Process 

Where, ft, it, and ot represent forget, input and output gates, respectively, and control the flow of 

information inside and outside the memory cells. This enables the LSTM network to model long-term 

temporal dependencies in time-series samples effectively. Simultaneously, CNN is employed for 

extracting spatial features from the multimodal data, such as but not restricted to image sequences and 

sensor arrays. Extracted features of CNN have been obtained by applying successive convolutions 

computed via equation 7, 

𝐹(𝑖, 𝑗) = ∑ ∑𝑋(𝑖 + 𝑚 − 1, 𝑗 + 𝑛 − 1) ⋅ 𝐾(𝑚, 𝑛) + 𝑏… (7)

𝑁

𝑛=1

𝑀

𝑚=1

 

Here, X(i,j) represents the input image data, 'K' the convolution kernel and 'b' the bias term. The 

convolutional layers in CNN capture the spatial patterns in the data, which are very much essential for 

accurate predictions in domains such as video analysis and sensor data fusion. The combination of 

output from XGBoost, LSTM and CNN through concatenation provides a fused feature representation 

given via equation 8, 

𝑍 = [𝑦’𝑋𝐺𝐵, ℎ𝑡, 𝐹(𝑖, 𝑗)]… (8) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

671 https://internationalpubls.com 

Where, y'XGB is the XGBoost prediction, ht is the LSTM hidden state and F(i,j) represents the CNN 

feature maps. This fused representation is further passed through fully connected layers for final 

prediction. To make the model more adaptive to dynamic environments proximal policy optimization 

is used. PPO optimizes the policy πθ(a∣s) in a way to balance the trade-off between exploration and 

exploitation. The objective function of PPO is presented via equation 9, 

𝐿𝑃𝑃𝑂(𝜃) = 𝐸’𝑡[𝑚𝑖 𝑛(𝑟𝑡(𝜃)𝐴’𝑡, 𝑐𝑙𝑖𝑝(𝑟𝑡(𝜃), 1 − 𝜖, 1 + 𝜖)𝐴’𝑡)] … (9) 

Where, is represented via equation 10, 

𝑟𝑡(𝜃) =
𝜋𝜃( 𝑎𝑡 ∣ 𝑠𝑡 )

𝜋𝜃𝑜𝑙𝑑( 𝑎𝑡 ∣ 𝑠𝑡 )
… (10) 

Which represents the probability ratio between the current and old policies, A't is the advantage 

estimate, and ϵ is a hyperparameter that controls the clipping range for stable updates. Therefore, the 

model will automatically change its parameters with dynamic updating to improve the prediction over 

changing environments. Finally, noise in the data is reduced by the KF-BNN, and there is uncertainty 

estimation of the prediction. The updated state estimates \\( x^t \\) and covariance Pt are given using 

the Kalman filter via equations 11 & 12, as follows : 

𝑥’𝑡 = 𝐴𝑥’(𝑡 − 1) + 𝐵𝑢(𝑡)… (11) 

𝑃𝑡 = 𝐴𝑃(𝑡 − 1) ∗ 𝐴𝑇 + 𝑄…(12) 

The Bayesian Neural Network then processes the filtered estimates, producing a distribution of 

possible outcomes P(y’∣D) based on the posterior probability represented via equation 13, 

𝑃( 𝑦’ ∣∣ 𝐷 ) = ∫ 𝑃( 𝑦’ ∣∣ 𝜃 )𝑃( 𝜃 ∣ 𝐷 )𝑑𝜃… (13) 

The model can therefore quantify the uncertainty in its predictions within this probabilistic framework 

and hence make the output more reliable, especially under noisy conditions. In summary, the model 

proposed in this paper leverages the complementary strengths of XGBoost, LSTM, CNN, and PPO to 

efficiently handle multiple modal data sample. These six equations reveal a more involved 

combination of feature importance, temporal dependencies, spatial patterns, real-time adaptability, and 

uncertainty quantification, whereby the features described have been robustly and flexibly endowed 

with the model in complex, dynamic data conditions. 

4. Comparative Result Analysis 

Experimental Setup: Experimental setup designed to analyze the performance of the proposed hybrid 

XGBoost-LSTM-CNN model with Proximal Policy Optimization in execution on several multimodal 

datasets comprising time-series, spatial, and tabular data samples. We evaluate the model's robustness 

in prediction with four datasets from heterogeneous domains: sensor data, financial time-series, 

medical data, and image-based spatial datasets & samples. The features that need to be extracted define 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

672 https://internationalpubls.com 

the way in which the different datasets are preprocessed: sequential dependencies, temporal trends, 

and spatial information sets. Each dataset was divided into training, validation, and test sets, 

respectively, using a 70/15/15 segregation in process. The results of the proposed model are compared 

with the three other established ones: [5], [8], and [14], which include state-of-the-art deep learning 

and hybrid methods for multimodal data fusion and/or time-series forecasting tasks. The four 

evaluation metrics-precision, accuracy, recall, and Mean Absolute Error-offered the best insight into 

the predictive performance and the model's capacity for handling multimodal data with good 

efficiency. Results will be provided in six tables, with thorough comparisons between the proposed 

model and methods [5], [8], and [14]. 

Table 1: Comparison on Sensor Dataset 

Method Precision (%) Accuracy (%) Recall (%) MAE 

Proposed Model 96.1 97.0 95.7 0.070 

Method [5] 93.0 95.2 92.5 0.090 

Method [8] 92.8 94.8 92.0 0.095 

Method [14] 91.5 93.7 90.2 0.102 

 

Figure 2. Comparison on Sensor Dataset Samples 

The results in figure 2 of the proposed model outperform all other metrics in the sensor dataset, with 

an accuracy of 96.1%, while bringing MAE down to 0.070. It improves the accuracy of method [5] by 

1.8% and for MAE reduces by 22%. Quite behind, the method in [14], for instance, has an MAE of 

0.102, which further illustrated the combined use of XGBoost and LSTM in noise reduction towards 

making better predictions. 

Table 2: Comparison on Financial Time-Series Dataset 

Method Precision (%) Accuracy (%) Recall (%) MAE 

Proposed Model 95.3 96.8 95.0 0.073 

Method [5] 93.7 95.3 93.2 0.085 

Method [8] 92.9 94.5 92.4 0.090 

Method [14] 91.2 93.9 91.0 0.098 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

673 https://internationalpubls.com 

The result of the proposed model outperforms the other methods, especially for the MAE with a value 

of 0.073, against the best results of the methods in [5] and [8] with 0.085 and 0.090 values, 

correspondingly. High precision and recall indicate much higher temporal dependency rate, captured 

by this model, comparing to all other works, having very important meaning for financial forecasting 

tasks. 

 
Figure 3. Comparison on MIMIC-III Healthcare Dataset Samples 

Table 3: Comparison on MIMIC-III Healthcare Dataset 

Method Precision (%) Accuracy (%) Recall (%) MAE 

Proposed Model 95.8 97.1 95.2 0.074 

Method [5] 94.0 95.5 93.7 0.085 

Method [8] 93.2 95.0 92.5 0.089 

Method [14] 92.0 94.3 91.8 0.095 

The performance of the proposed model using the MIMIC-III dataset on patient time-series data 

indicates high accuracy at 97.1%, MAE reduced by about 13% compared to method [5]. In the 

proposed model, the use of Bayesian Neural Networks and Kalman filtering helps reduce noise from 

healthcare sensor data; hence, reliable predictions are achieved in the process. 

Table 4: Comparison on Image-Spatial Dataset 

Method Precision (%) Accuracy (%) Recall (%) MAE 

Proposed Model 94.6 96.2 94.0 0.078 

Method [5] 92.5 94.5 92.0 0.091 

Method [8] 91.8 94.0 91.2 0.093 

Method [14] 90.9 93.5 90.0 0.100 

Having considered the precision, on spatial data tasks like image processing, the proposed model 

achieved a precision value of 94.6% with an MAE of 0.078, outperforming method [5] by 14% error 

reduction. This was contributed by the capability of the CNN component in capturing the features on 

a spatial context while the hybrid fusion of XGBoost, LSTM, and CNN made the proposed model 

highly effective for the process. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

674 https://internationalpubls.com 

Table 5: Real-Time Adaptation with PPO on Sensor Dataset 

Method Precision (%) Accuracy (%) Recall (%) MAE 

Proposed Model (with PPO) 96.5 97.6 96.1 0.072 

Proposed Model (without PPO) 96.1 97.0 95.7 0.070 

Method [5] 93.0 95.2 92.5 0.090 

Method [8] 92.8 94.8 92.0 0.095 

The proposed model with the turned-on PPO has some improvements on precision, accuracy, and 

recall, and the MAE gets reduced compared to the no-PPO settings. The presence of PPO is underlined, 

which means that in a dynamic environment, things need real-time adjustments towards a best 

performance in different scenarios. 

Table 6: Federated Learning Comparison on Distributed Healthcare Dataset 

Method Precision 

(%) 

Accuracy 

(%) 

Recall 

(%) 

MAE 

Proposed Model (with Federated Learning) 94.9 96.5 94.5 0.082 

Proposed Model (without Federated 

Learning) 

94.6 96.2 94.0 0.084 

Method [5] 93.5 95.3 93.2 0.090 

Method [8] 92.9 94.5 92.4 0.092 

Federated learning proves useful in distributed environments, such as in healthcare systems when data 

privacy plays a major role. Because the local models contribute to the much more generalized global 

model without the loss of any sensitive data samples of patients, the improvement in the proposed 

model with the help of federated learning is enhanced by 0.3% in precision and reduced MAE by 2%. 

These results establish the fact that the proposed XGBoost-LSTM-CNN hybrid model integrated with 

PPO, Kalman Filter-enhanced Bayesian Neural Networks, and Federated Learning significantly 

outperforms existing methods on a wide range of multimodal datasets. These improvements in the 

values of precision, accuracy, and MAE for various tasks prove the strength of the model in terms of 

versatility and robustness. 

5. Conclusion & Future Scopes 

Such a hybrid model will be able to handle this complex, multi-modal dataset much better. The 

proposed XGBoost-LSTM-CNN hybrid model with Proximal Policy Optimization makes huge gains 

in precision, accuracy, while cutting down error. It combines the robust tabular data handling capability 

of XGBoost with the ability of LSTM to model long-term temporal dependencies and CNN spatial 

feature extraction. Therefore, the proposed model mitigates major shortcomings of a number of 

prevailing techniques for multimodal fusion in such a way. This model is more adaptable to dynamic 

environments, since it is combined with Proximal Policy Optimization where real-time model updates 

can be performed in a continuous stream of changing data. These are verified through experimental 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

675 https://internationalpubls.com 

results on various sensor, financial, healthcare, and image-spatial datasets. For example, the sensor 

dataset tested on the proposed model had an accuracy of 97.0% and a MAE of 0.070, which was 1.8% 

better in terms of accuracy than method [5] with a reduction in MAE by 22%. In the Financial time-

series dataset, the model has achieved 96.8% accuracy and MAE 0.073, which shows a major gain 

over methods [8] and [14]. The model provided uncertainty estimation along with noise reduction, 

incorporating Kalman Filter-enhanced Bayesian Neural Networks on healthcare applications. It 

achieved an accuracy of 97.1% on the MIMIC-III dataset, with a 13% reduction in MAE compared to 

that from method 5. It also showed the adaptability of the proposed model using PPO in dynamic 

tuning, increasing precision by 0.4% and reduced MAE to 0.072 in the case of real-time sensor data. 

Federated learning contributed much to the effectiveness of the model in a privacy-preserving 

environment for widely distributed healthcare applications with an accuracy of 96.5% and reduced 

MAE to 0.082. These findings hint at the robustness and scalability of the proposed model in analyzing 

multimodal data to achieve further improvements of predictive accuracy and reliability in a number of 

domains. 

Future Scope of Work 

Although the proposed model has shown quite a high degree of accuracy and performance in 

multimodal datasets, further avenues of research and enhancement are possible. Two of the most 

promising directions in which further development could take place are: a) extension of the current 

architecture to higher-dimensional and more complex multimodal data, such as 3D spatial information 

from medical imaging or higher-resolution temporal video streams in real-time instance sets. Attention 

mechanisms, again developed around transformers, can be combined to selectively pay more attention 

to most salient features across all modalities, hence further improving predictive accuracy. Another 

improvement can be done at the optimization of the federated learning framework. Added mechanisms 

for differential privacy could enhance guarantees on data privacy without losing model performance 

in healthcare and financial applications. Blockchain-based federated learning methods could also 

consider a well-decentralized training process and further improve the robustness and scalability of the 

model against distributed environments. In this respect, another exciting direction would come with 

endowing the model with the ability to learn-to-adapt quickly to new tasks with minimal retraining by 

integrating meta-learning techniques into the model. That would serve very useful in dynamic 

environments where data distributions are constantly shifting, say in autonomous systems or financial 

markets. Finally, it would also be worth including in the Bayesian Neural Network module some of 

the advanced probabilistic reasoning methods that have been demonstrated for better uncertainty 

quantification, such as variational inference levels. It would make the model more robust against noisy 

data and therefore would permit better outlier detection. In all, the proposed hybrid architecture opens 

a very promising avenue for further work process. There is considerable room for refinement and 

extension of its capabilities in undertaking multi-modal data-fusion tasks on scales of increasing 

complexity, in real-time, and larger magnitudes. 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

676 https://internationalpubls.com 

6. References 

[1] L. Junyong, T. Yinyin, Z. Delin, Y. Feifei and Z. Yufeng, "Fault Prediction of Electromagnetic 

Launch System Based on Knowledge Prediction Time Series," in IEEE Transactions on Industry 

Applications, vol. 57, no. 2, pp. 1830-1839, March-April 2021, doi: 10.1109/TIA.2020.3046705. 

keywords: {Feature extraction;Time series analysis;Electromagnetics;Prediction algorithms;Expert 

systems;Electromagnetic scattering;Railguns;Electromagnetic launch (EML) system;expert 

system;fault prediction;health monitoring;neural network;time series prediction}, 

[2] D. -K. Kim and K. Kim, "A Convolutional Transformer Model for Multivariate Time Series 

Prediction," in IEEE Access, vol. 10, pp. 101319-101329, 2022, doi: 

10.1109/ACCESS.2022.3203416. 

keywords: {Time series analysis;Predictive models;Convolutional neural networks;Data 

models;Forecasting;Transformers;Feature extraction;Artificial neural networks;predictive 

models;time series prediction}, 

[3] D. Zeng, J. Lu and Y. Zheng, "Combined Fuzzy Time Series Prediction Method for Fault 

Prediction of EML Pulse Capacitors," in IEEE Transactions on Plasma Science, vol. 49, no. 2, pp. 

905-913, Feb. 2021, doi: 10.1109/TPS.2020.3029840. 

keywords: {Capacitors;Capacitance;Time series analysis;Discharges 

(electric);Electromagnetics;Degradation;Uncertainty;Electromagnetic launch (EML) 

system;fuzzy time series prediction;large-scale pulse forming network (PFN);metallized film 

capacitors}, 

[4] S. Feng, M. Han, J. Zhang, T. Qiu and W. Ren, "Learning Both Dynamic-Shared and Dynamic-

Specific Patterns for Chaotic Time-Series Prediction," in IEEE Transactions on Cybernetics, vol. 

52, no. 6, pp. 4115-4125, June 2022, doi: 10.1109/TCYB.2020.3017736. 

keywords: {Task analysis;Time series analysis;Dynamical systems;Predictive models;Market 

research;Optimization;Heuristic algorithms;Dynamic pattern;multitask learning (MTL);Stiefel 

manifold optimization;time-series prediction}, 

[5] F. Shen, J. Liu and K. Wu, "Multivariate Time Series Forecasting Based on Elastic Net and High-

Order Fuzzy Cognitive Maps: A Case Study on Human Action Prediction Through EEG Signals," 

in IEEE Transactions on Fuzzy Systems, vol. 29, no. 8, pp. 2336-2348, Aug. 2021, doi: 

10.1109/TFUZZ.2020.2998513. 

keywords: {Time series analysis;Electroencephalography;Forecasting;Prediction algorithms;Brain 

modeling;Predictive models;Proposals;1D-convolutional neural network (1D-CNN);elastic 

net;electroencephalogram (EEG);fuzzy cognitive map (FCM);time series prediction}, 

[6] X. Zhou, N. Zhai, S. Li and H. Shi, "Time Series Prediction Method of Industrial Process With 

Limited Data Based on Transfer Learning," in IEEE Transactions on Industrial Informatics, vol. 

19, no. 5, pp. 6872-6882, May 2023, doi: 10.1109/TII.2022.3191980. 

keywords: {Time series analysis;Predictive models;Data models;Transfer learning;Neural 

networks;Adaptation models;Production;Industrial time series prediction;multistep 

prediction;multitask learning;transfer learning}, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

677 https://internationalpubls.com 

[7] X. Yin, Y. Han, H. Sun, Z. Xu, H. Yu and X. Duan, "Multiple Attention Generative Adversarial 

Network for Multivariate Time Series Prediction," in IEEE Access, vol. 9, pp. 57351-57363, 2021, 

doi: 10.1109/ACCESS.2021.3065969. 

keywords: {Time series analysis;Predictive models;Autoregressive processes;Generative adversarial 

networks;Data models;Market research;Correlation;Multivariate data;time series 

prediction;multiple attention;generative adversarial network}, 

[8] J. Yi, J. Huang, W. Zhou, G. Chen and M. Zhao, "Intergroup Cascade Broad Learning System 

With Optimized Parameters for Chaotic Time Series Prediction," in IEEE Transactions on 

Artificial Intelligence, vol. 3, no. 5, pp. 709-721, Oct. 2022, doi: 10.1109/TAI.2022.3143079. 

keywords: {Time series analysis;Predictive models;Learning systems;Algorithm design and 

analysis;Broad learning system (BLS);chaotic time series prediction;multiobjective;multiple 

model}, 

[9] X. Chen and L. Sun, "Bayesian Temporal Factorization for Multidimensional Time Series 

Prediction," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, 

pp. 4659-4673, 1 Sept. 2022, doi: 10.1109/TPAMI.2021.3066551. 

keywords: {Time series analysis;Data models;Bayes methods;Spatiotemporal 

phenomena;Tensors;Reactive power;Probabilistic logic;Time series prediction;missing data 

imputation;low rank;matrix/tensor factorization;vector autoregression (VAR);Bayesian 

inference;Markov chain Monte Carlo (MCMC)}, 

[10] F. Mubang and L. O. Hall, "VAM: An End-to-End Simulator for Time Series Regression and 

Temporal Link Prediction in Social Media Networks," in IEEE Transactions on Computational 

Social Systems, vol. 10, no. 4, pp. 1479-1490, Aug. 2023, doi: 10.1109/TCSS.2022.3180586. 

keywords: {Social networking (online);Time series analysis;Predictive models;Blogs;Task 

analysis;Software development management;Computational modeling;Extreme gradient 

boosting;link prediction;social media;time series prediction}, 

[11] C. Ma, G. Dai and J. Zhou, "Short-Term Traffic Flow Prediction for Urban Road Sections Based 

on Time Series Analysis and LSTM_BILSTM Method," in IEEE Transactions on Intelligent 

Transportation Systems, vol. 23, no. 6, pp. 5615-5624, June 2022, doi: 

10.1109/TITS.2021.3055258. 

keywords: {Time series analysis;Predictive models;Fractals;Data models;Correlation;Biological 

neural networks;Training;Traffic engineering;short-term traffic flow prediction;LSTM_BILSTM 

method;time series analysis;urban road section}, 

[12] W. Wang, W. Liu and H. Chen, "Information Granules-Based BP Neural Network for Long-Term 

Prediction of Time Series," in IEEE Transactions on Fuzzy Systems, vol. 29, no. 10, pp. 2975-

2987, Oct. 2021, doi: 10.1109/TFUZZ.2020.3009764. 

keywords: {Time series analysis;Market research;Predictive models;Neural 

networks;Forecasting;Hidden Markov models;Semantics;Back-propagation neural 

network;information granulation;long-term prediction;time series forecasting}, 

[13] N. Maaroufi, M. Najib and M. Bakhouya, "Predicting the Future is Like Completing a Painting: 

Towards a Novel Method for Time-Series Forecasting," in IEEE Access, vol. 9, pp. 119918-

119938, 2021, doi: 10.1109/ACCESS.2021.3101718. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 2 (2025) 

 

678 https://internationalpubls.com 

keywords: {Forecasting;Predictive models;Time series 

analysis;Tools;Testing;Extrapolation;Analytical models;Scientific prediction and experimental 

philosophy of science;extensive structural realism;bridging philosophy;time series 

forecasting;fully integrated modeling and processing framework;ensemble data autocorrelation 

forecasting;augmented dimension prediction;image and signal processing}, 

[14] S. Ren, B. Guo, K. Li, Q. Wang, Z. Yu and L. Cao, "CoupledMUTS: Coupled Multivariate Utility 

Time-Series Representation and Prediction," in IEEE Internet of Things Journal, vol. 9, no. 22, 

pp. 22972-22982, 15 Nov.15, 2022, doi: 10.1109/JIOT.2022.3185010. 

keywords: {Couplings;Time series analysis;Sensors;Internet of Things;Correlation;Predictive 

models;Tensors;Coupling relational learning;multivariate utility time series (MUTS);sensory data 

modeling;smart cities;utility demand prediction}, 

[15] K. Yuan, K. Wu and J. Liu, "Is Single Enough? A Joint Spatiotemporal Feature Learning 

Framework for Multivariate Time Series Prediction," in IEEE Transactions on Neural Networks 

and Learning Systems, vol. 35, no. 4, pp. 4985-4998, April 2024, doi: 

10.1109/TNNLS.2022.3216107. 

keywords: {Time series analysis;Feature extraction;Spatiotemporal 

phenomena;Correlation;Representation learning;Predictive models;Prediction algorithms;Fuzzy 

cognitive maps (FCMs);fuzzy neural network;multivariate time series prediction (TSP);sparse 

autoencoder (SAE);spatiotemporal features},