Intelligent Modelling of Magnetic Hyperthermia Parameters for Predicting Anti-Tumor Effects

Jayashree Rajesh Prasad^1*, Rajesh S. Prasad¹ and Nanasaheb D. Thorat^{2 *}Correspondence: Jayashree Rajesh Prasad, Department of Biotechnology, MIT ADT School of Computing, MIT ADT Technology University, Pune, India, Email:

Author info »

Introduction

The use of energy-absorbing nanoparticles (hyperthermia or heat mediated therapy) has become increasingly important (Dar MS, et al., 2021; Ciaburro G and Iannace G, 2021). Through Magnetic Nanoparticles (MNPs), the malignant area containing the ther mos seeds is heated to a high degree. Hysteresis, Néel and Brownian relaxation losses cause heating in magnetic nanoparticles when they are exposed to an alternating magnetic field. Brown relaxation occurs when nanoparticles rotate in fluid and generate heat by a fraction process in the aqueous medium. Heat dissipaion to the surrounding area is limited due to reduced blood flow in the tumor area, which has disordered blood vessels.

Since the turn of the century, Artificial Intelligence (AI) has dominated health care and medical science. Researchers have decided that combining this technology with doctors can be beneficial, however AI’s incredible potential to transform health care into a far more modern and efficient system has yet to be studied. In this evidence, we have also used AI to create better nanoparticles for hyperthermia treatment.

Furthermore, machine learning is a useful tool for studying and discovering latest developments from time series data. Because of the non-linear behavior of our time series, we picked an artificial neural network for mathematical modeling because they can do non-linear mappings. In the field of time series forecasting, this technique has proven to be extremely useful.

The goal of this study is to investigate what makes magnetic NFs effective heating agents in magnetic hyperthermia. We propose using artificial intelligence, machine learning, and deep learning to evaluate the impact of various polyol synthesis parameters on the size, shape, chemical composition, number of cores, and crystallinity of the final NFs. In Magnetic Hyperthermia (MHT) and Photo Thermal (PT) therapy, these nano characteristics are connected to the NFs’ magnetic, optical, and electrical properties, as well as their aggregate macroscopic thermal properties, to discover correlations between their structure, attributes, and function (Figure 1).

Figure 1: Pictorial representation of the proposed idea

We foresee another goal of AI/machine learning based modeling of this issue as well. Hyperthermia is a way of eliminating tumor cells by heat. The heating device is frequently an ultrasonic transducer. The temperature distribution generated by the ultrasound must be planned in order to kill tumor cells while not harming normal tissue. The phase and amplitude of the input signal for each element of a multi-element ultrasound transducer can be modified to provide a proper temperature distribution to match the needs of particular treatments. Direct computation, on the other hand, can be time-consuming, and determining the ultrasound transducer parameters with a particular temperature distribution might be tricky. ANNs are utilized in this proposal to learn the association between ultrasound transducer parameters and temperature distribution, both in the laboratory and in the field.

The input data in this case is time-series data. Machine learning-based algorithms gain their expertise on their own, based on the data patterns they receive, rather than requiring particular beginning inputs from the creator (Ciaburro G, 2020). In machine learning models, the machine can determine the patterns to follow on its own (Figure 2).

Figure 2: Machine-learning-based knowledge extraction for time series is depicted as a block diagram

As a result, autonomy is the true differentiator between artificial intelligence and human intelligence. The system receives feedback during the learning process that separates these techniques. The set of data required for training, with the goal of calculating the associations between input data, output data and the parameters of the model estimated by the system are represented.

In these circumstances, we refer to supervised learning (Caruana R and Niculescu-Mizil A, 2006), in which the input data is given a label that represents the model’s output, and the system learns to detect the links between the input data and the labels. In other cases, the data is not labeled, which is referred to as unsupervised learning (Celebi ME and Aydin K, 2016), and the system only receives the input datasets with no indication of the outputs that will be obtained. The goal of this learning method is to evaluate a logical structure in the incoming data that hasn’t been categorized before. After obtaining the input data, the computer must achieve a goal that is determined by a value function that only tells it if the goal has been met. In the subsections that follow, we’ll look at the most common machine learning methods that the scientific community has used to handle a variety of challenges related to time series forecasting.

Methods to gather data

To accomplish so, we look at the MHT efficiency, as well as the NFs’ optical responses and collective photo thermal properties. Since the turn of the century, AI has dominated health care and medical science. Our partnership with doctors has the potential to determine the ability of AI to transform health care into a far more modern and efficient system. Thus, we present AI models for estimating the production of better nanoparticles for hyperthermia therapy. We also offer a deep learning approach for hyperthermia data time series modeling. ANNs with external input must extract information from previous values in the data and analyze it in order to learn and improve. We envisage a simple and quick method for producing iron oxide nanostructures with superior heating properties. This technique will examine the polyol process and characteristics such as precursor concentration, polyol molecular weight, and reaction time, with the goal of producing NPs with the fastest heating rates feasible. The visualization based on Deep Neural Networks (DNNs) intends to investigate the impacts of Polyacrylic acid during the creation of outstanding nano heating agents such as Iron Oxide Nano flowers (IONFs). This study will also look into the impact of using a seeded growth approach to gradually expand the size of the NFs.

Literature Review

Technological background

Modeling task can be feasibly performed with multi-layer neural networks, which have the advantage of real-time computation after sufficient training. Deep learning models are known to achieve exceptional performance while maintaining a level of precision that is acceptable. Although these models have not produced exact reference values as seen by many researches, the temperature fields produced are extremely similar to those of the reference. The ability to change treatment planning in real time is enabled by the quick computation of the temperature field in hyperthermia i.e., dynamic therapy planning, as we call it. For time series modeling of hyperthermia data, deep learning method has been explored. Deep learning is a subset of machine learning, which itself is a part of a larger area known as artificial intelligence. Artificial Neural Networks (ANN)-based algorithms will make up deep learning. Various studies show that scientists have employed Artificial Neural Networks (ANNs) to analyze time series data for prediction and forecasting.

To determine the optimal nanoparticle, we plan to create a system using six artificial neural networks. ANNs with exogenous input have the ability to collect information from past data values and process it to learn and then anticipate step forward values, giving us an understanding of each particle’s performance. However for better predictions, the algorithm would require as much data as feasible.

Implementation

We intend to use artificial neural network modeling to examine localized anticancer effects. The optimal nano hybrid composition will be determined using a neural net time-series model. Each of the components will be represented by one of six Nonlinear Autoregressive with External Input (NARX) models. The accuracy of the projected results will be evaluated using the Mean Squared Error (MSE). In the training phase, the nano hybrid containing 45 percent magnetite and 55 percent graphene (F45G55) has the greatest MSE value of 0.44703, which is where the model may attain optimal results after considerable epochs. With the maximum Specific Absorption Rate (SAR) and MSE values, the F45G55 nano hybrid is projected to be the best for hyperthermia applications in low dosage.

After a long period of research, the scientific community has made significant progress in the field of magnetic hyperthermia over the previous two decades. All of the advancements in issues ranging from nanoparticle manufacturing to biocompatibility and in vivo testing have been aimed towards accelerating the development of new clinical trials. They did not move at the expected rate, as many others did. Today, collaboration and perception gained from a rigorous examination of the lessons learnt from seminal clinical studies enable us to have a better world.

A more definitive take-off of this actual cancer nano therapy, the future seems to be better. This opinion evaluation presents a combination of stateof-the-art, deliberately highlighting innumerable complications.

We have identified and explored the following ANN and DNN models that could be applied to implement the problem at hand. We are in the process of review and evaluation of these models. However, we have elaborated an approach in the technology section of this proposal; we will explore the following models too.

• Supervised Multilayer Perceptron (MLP) (Azadeh A, et al., 2007)

• The ANN model through the analysis of variance with Analysis of Variance (ANOVA) technology (Miller Jr RG, 1997).

• ANN and the conventional regression model (Hill T, et al., 1996)

• Makridakis competition, called M-Competition (Makridakis S, et al., 1982)

• Hybrid methodology that combines Autoregressive Integrated Moving Average (ARIMA) (Zhang GP, 2003) and ANN models (Contreras J, et al., 2003)

• Hybrid methodology capable of exploiting the strengths of traditional time series (Jain A and Kumar AM, 2007)

• Hybrid approach, this time by exploiting the seasonal adjustment potential of the SARIMA model (Tseng FM, et al., 2002; Chen CF, et al., 2009)

• Hybrid model based on the integrated self-regressive moving average algorithm (ARIMA) and artificial neural networks (Khashei M and Bijari M, 2010).

• Multilayer feed forward neural network for forecasting the exchange rate in a multivariate framework (Chaudhuri TD and Ghosh I, 2016)

• Strategy for selecting an ANN-based forecasting model (Aras S and Kocakoç İD, 2016)

• Forecasting method for renewable energy sources to achieve intelligent management of a micro grid system (Doucoure B, et al., 2016)

• Time series of dynamic light dispersion with the use of ANNs (Chicea D and Rei SM, 2018).

• Method of exponential time series alignment with time series alignment using artificial neural networks (Horák J and Krulicky T, 2019)

• Three hybrid models to predict wind speed (Liu H, et al., 2013).

• Neural network time series analyses (Wang CC, et al., 2010).

Local community need

Social needs of a cancer patient’s community: Many requirements that are deemed to be social in character appear to be linked to an individual’s health. It is incredibly painful for families to come across their loved ones encounter problems due to health which causes inability to manage their chores and health or might require 24-hour care. Given that differences in tumor biology and genetic variants have not completely explained the persistent Black/White breast cancer mortality disparity, more attention to social factors such as race/ethnicity, socioeconomic status, and others is needed across the cancer continuum, including breast cancer.

Emotional needs of cancer patient’s community:

• Scars, weight fluctuations, the loss of a breast or other body part, or other changes to the body may be affecting the body image and self-esteem. Some elements of the affected body may not function as well as they once did. It is important to allow the affected patient to grieve for the things that they have lost. Thus, support groups, counselling, exercise, and a well-balanced diet can all be beneficial.

• According to the National Cancer Institute, depression affects around two out of every ten cancer patients. It’s critical to speak with caretaker about seeking treatment if someone is experiencing persistent sadness or no longer enjoy their favorite activities.

• Fear of cancer recurrence which is natural to be concerned that every ache or sniffie could signify a recurrence of cancer, especially in the first year after treatment. Medical appointments, anniversaries, and other events can elicit or exacerbate these anxieties. Accepting anxiety, managing stress, and focusing on things, following-up appointments, can be beneficial.

• Finding meaning-Many cancer sufferers discover purpose and fresh starts in their lives over time. Some people believe they are stronger or more capable. Some people are motivated to do new activities, while others discover that they value each day more. New insights, on the other hand, can take years to develop and may not show right away.

• Grief-It is natural, even expected, to feel bad for the loss of healthy life. Many patients, however, are taken aback by the intensity of their feelings about the need to transition to a new normal. Feelings like regret, guilt and being upset for losses can be overwhelming. Thus such patients require healthcare professional.

• Guilt-Knowing that some patients did not survive makes the affected person feel guilty. Some people are concerned that they are putting undue strain on caretakers and family members. Counselling or a support group can help the patient express their emotions.

• Loneliness-Feeling of loneliness is frequent after a cancer diagnosis. Thoughts of belief that people are incapable of fully comprehending the affected person’s situation might cause loneliness, for which counselling, may provide emotional assistance.

• Relationships-Cancer can be a burden on friendships, family, and co-workers. People may be unsure of the effected patient’s thoughts. Support from friends and family members would be a therapy.

• Spirituality-As a result of cancer treatment, one may develop a new perspective on conviction. Traditions are renewed by some survivors who have found new connections and community.

• Anxiety can be exacerbated by the changes that come with illness, treatment, and survivorship. Recurrence of issues may be experienced during the treatment. Proper guidance, timely exercise, social interaction, relaxation techniques, meditation, and artistic pursuits might be beneficial

• Re-entering social and professional life after cancer can be difficult. Many people are concerned about an increased risk of illness, a lack of energy, and work performance worry. Work, on the other hand, can provide a sense of normalcy. Open communication with co-workers can help to remove the feelings of insecurity, but, information related to treatment for choice for cancer is to be properly monitored.

• Medicinal treatment needs-Hyperthermia, or the dealing of ailments using heat, has been around since the establishment of time. Increasing the temperature of cells above 41°C is known to have a number of effects on the cell membrane and interior, including (a) increasing the fluidity and permeability of the cell membrane, (b) slowing down the mechanisms of nucleic acid and protein synthesis, (c) inducing protein denaturation and agglomeration, and (d) damaging the tumor vasculature, resulting in a reduction in blood flow.

Technological needs

As a result, this is a demanding multidisciplinary research field where insights are required to study, which are listed below. The principles of heat generation by MNPs in AMF, the limitations and challenges of MH, and the applications of MH employing multifunctional hybrid MNPs are all scoped in this research. Through modeling of parameters using ANNs or DNNs, this research will establish ideal associations of various parameters dealing with hyperthermia. This paper addresses and justifies clarification of the following aspects-

• Identifying the biocompatible, biodegradable MNPs that have good colloidal stability

• Successful penetration of such MNPs to tumor cells

• Optimized use of MNPs which necessitates MNPs with large heat generation capabilities

• Identification of an optimized or limited frequency and amplitude ranges for the AMF meeting safety conditions used to heat the MNPs

• Tracking of accurate temperature for a non-invasive approach at the cellular level

• Identifying the factors impacting heat transfer from MNPs to cells

• Thorough understanding the influence of temperature on the biological mechanisms of cells

Importance of local community needs

According to a new data, cancer incidence in India climbed at an annual pace of 1.1%-2% on average from 2010 to 2019. The number of cancer deaths in the country increased by 0.1%-1% on average over the same time period, according to the study. According to a report published in JAMA on December 30, 2021 by the Institute for Health Metrics and Evaluation (IHME), University of Washington School of Medicine, the growth rates are among the highest in the world.

According to forecasts from another study, 1,392,179 persons in India will be diagnosed with cancer by 2020. Researchers discovered that the breast, lung, mouth, cervix, uterus, and tongue were the five most common sites of the disease.

According to the Cancer Statistics Report, 2020, the expected incidence for men was 94.1 per 100,000 people, while for women it was 103.6 per 100,000 people. Apart from these limitations, there is a need to improve the effects of radiotherapy and chemotherapy, as conventional chemotherapy has the following side effects. Chemotherapy has a number of side effects that can cause a lot of pain for the patient. This should be discussed with health professional. These symptoms can be controlled with a variety of drugs. The treatment may or may not enhance the cancer’s overall prognosis or lengthen a person’s life expectancy.

When receiving palliative chemotherapy, one’s quality of life may suffer, and if the treatment is unlikely to improve one’s life expectancy, one must consider if the treatment is worthwhile.

Some chemotherapy regimens can result in death in a small percentage of patients, and it is important to be aware of these serious implications before beginning treatment.

Challenges

Recent computing experiments explicitly state the need for development of Artificial Intelligence (AI) and Machine Learning (ML) algorithms, which will accelerate the discovery and design of new magnetic materials. There are following challenges to meet this need.

• Data-storage technologies depend on materials that sustain magnetic properties at high temperature. While researchers have a range of such materials to work with, theory suggests that the known options are but a small fraction of the high-temperature magnets that are possible. To speed up the discovery and design of new high-temperature magnets, there is an immense need to develop several machine-learning models that can predict the temperature at which a material demagnetizes-its Curie temperature-from its chemical composition.

• For an empirical data from 2500 known ferrimagnets, there is a requirement to analyze the predictive models and evaluate their accuracy. Each model describes the correlation between a material’s Curie temperature and several other properties, such as its atomic number, its melting temperature, and the type of bonds that form between the atoms.

• To distinguish between polymorphs through machine learning/artificial intelligence. For some identical materials whose Curie temperatures differ because of their distinct structures.

• The AIML models has some fundamental error that cannot be reduced by increasing the amount of training data, and means that, although the algorithm can accelerate the materials design process, researchers need other methods to confirm the predicted properties.

Discussion

Neural computation is based on the structure of biological neural networks, which it attempts to mimic in order to simulate their essential operations. Inspired by biological systems, an Artificial Neural Network (ANN) considers a large number of processors known as artificial neuron (Abiodun OI, et al., 2018). The brain is a nonlinear, highly complicated system that can process information in parallel. In this way, it can execute some jobs faster than any current computer, such as pattern recognition and movement management, but most importantly, it can learn from prior experience. Our model based on artificial neural networks is thus made up of neurons that can convey signals to neighboring neurons via connections (Figure 3). We can distinguish between the following in this structure-Neurons that receive information from the environment are known as input neurons. Hidden neurons: Input neurons use them to communicate with output neurons or other hidden neurons; Output neurons: They respond to the environment by emitting replies. The transmission of messages between neurons when the signal they receive surpasses a specific activation threshold is the process that regulates the operation of such a system.

Figure 3: Flow of input to output with illustration of processing parameters

y_i=f(x_i)                    (1)

As a result, the activity of an artificial neuron can be summed using the mathematical function defined by the following equation-

The activation function is a nonlinear function that is typically nonlinear. Equation (3) represents the term xi in equation-

X_i=Ʃw_ji+b_i               (2)

Where, w_ji=connection weights; and bi=bias. As they summarize the information extracted during the learning process, the parameters given in both the equations are critical for the effectiveness of neural network investigations. The parameters are updated during the learning process to get closer to more accurate and precise prediction models (Da Silva IN, et al., 2017). For time series data, an artificial neural network design is used. The input dataset is separated into groups with a predetermined number of components. Each of these vectors is a separate object.

Figure 4 shows the construction of an artificial neural network handling time series data. The input dataset is separated into groups with a predetermined number of components. Each of these vectors is a network input that is matched with the correct output. Figure 4 depicts the construction of a simple feedforward artificial neural network. Unidirectional connections between neurons belonging to immediately adjacent layers are envisioned in feed-forward networks, with the link moving forward, from layer n to layer n+1. As a result, in this architecture, the inputs are only connected to the neurons of the hidden layer (of which there may be more than one), which are then connected to the neurons of the output layer.

Figure 4: A typical neural network architecture for time-series data

The weight assumed by the connections is weighted using a simple linear combination of the input variables in equation (2). It’s obvious that a simple structure like this can’t model complicated systems with nonlinear features. By processing the result through a nonlinear function called the activation function, this simple discriminant function can be generalized. When the input neurons receive a signal from the outside that is above a threshold and propagate it towards the inside of the network, the activation function conditions the interaction between the input and output neurons. The neuron is activated and participates in the processing of the model output when the activation threshold is achieved (Alanis AY, et al., 2019).

Otherwise, the neuron remains dormant and does not contribute to the output signal’s creation. Individual activation functions are distinguished depending on the structure employed to represent the link between the different neurons.

The activation function is a nonlinear function that is both continuous and differentiable. Nonlinear because nonlinearity features are required for a network to accomplish a complicated mapping of the input information. The features of continuity and differentiation, on the other hand, are required for the error’s retro-propagation (Romero VP, et al., 2016).

The weights must be modified when the activation function of the neurons and the design of the network have been suitably chosen so that the network functions optimally. Supervised learning is one approach for training a neural network, in which the network is given a set of inputs, or training set, and is asked to understand how the phenomena evolves over time in order to estimate the parameters of the model (Khosrow-Pour DB, 2018) (Figure 5).

Figure 5: A typical neural network training process

The calculated error is then applied to the weights after the response achieved with these inputs is compared to the observed response. In this method, the best suitable model may be tested on a sample, the test set that was not utilized in the training phase to assess the neural network’s generalization capacity (Al-Massri R, et al., 2018). The training phase will have to be repeated if this network is unable to give appropriate outputs for inputs that have never been provided.

Basic computational capability linked to additional units of the same kind Machine learning is a complementary tool for analyzing the hidden developments in time series. Due to non-linear behavior of our time series, artificial neural network for mathematical modeling has been chosen as they possess the ability to carry out non-linear mappings. This technique has been of considerable usability in the field of time series forecasting. It is required to divide the input data in advance when using an artificial neural network model for time series prediction (Wang L, et al., 2018). The method to use is to partition the data set into sequential groups with a predetermined number of components.

Assume we have a distribution of univariate time series data, and the explanatory variable of the phenomenon of interest is denoted by xi with data I=1,2,...,n.

Several nonlinear autoregressive structure-based models have also been suggested (Qi M, Zhang GP, 2001; Khashei M and Bijari M, 2010; Tealab A, et al., 2017). Using the neural net time series programme, we plan to create a neural net time-series model. For prediction, one or more-time series are required; however, our time-series challenge is to create a prediction using the past values of the predicted time series (y(t)) and another time series (x(t)), in this case the 100 percent magnetite particle. We will employ a Nonlinear Autoregressive with External Input (NARX) neural network as a foundation.

NARX is a recurrent network that can represent dynamic systems. It can not only predict output values that are regressed on previous values, but it can also be used for nonlinear filtering After testing with different combinations, the size of the hidden layer will be carefully adjusted, and time delays will be included to incorporate the dynamic of the input dataset.

The time steps will be divided initially into three parts-70% for training, 15% for validation, and 15% for testing. The network will also be trained using the Bayesian Regularization (BR) back propagation algorithm. To solve the overfitting problem, regularization is required. Generalization mistakes could be utilized deftly to combat the performance degradation. There are a variety of approaches that can be used to accomplish this (Goodfellow I, et al., 2016). Our choice of Bayesian regularization, on the other hand, is influenced by the Bayesian theorem, which combines prior data with the maximum likelihood function to get a posterior distribution.

Due to its ability to reveal complicated data patterns and interrelations (Medeiros MC and Pedreira CE, 2001), BR is a superior choice for quantitative research (Kayri M, 2016). The most reliable strategy for solving a problem like future prediction that is based on uncertainty is to use a probabilistic approach to machine learning (Murphy KP, 2012). The method is a complex, but it produces better results and has proven to be the most resilient and vigorous of the back propagation NNs. Regularization is designed to avoid overtraining and overturning, hence models trained with BR are difficult to over train and overturning (Burden F and Winkler D, 2009). It makes use of posterior probability, which entails using the Bayesian theorem for parametric optimization and updating prior information to posterior knowledge (Broemeling LD, 2017).

Bayesian inference is a fantastic method for statistically analyzing stochastic systems (Akram KB, et al., 2020). The training procedure is repeated until the best result is obtained, and then it is stopped when the generalization stops improving; the model can be trained for a maximum of 1000 epochs. Following training, the accuracy of anticipated results is assessed using Mean Squared Error (MSE), which is the average of the squared difference between output and goal values. Furthermore, Regression (R) values were generated to determine the association between output and target.

Contribution to the society

The quality of patient outcomes in a hospital setting is deteriorating. It is critical to establish a paradigm change for better patient outcomes. A traditional strategy to a machine-driven approach are given below-

• The clinician workforce is insufficient to cater to the growing population with varying conditions.

• Scaling up physicians is difficult, while scaling up machines is simple.

• In private hospitals, the consumerist approach has an impact on the quality of patient outcomes.

• A bureaucratic bottleneck, particularly in government-run hospitals, which leads to poor health outcomes.

A machine-driven strategy would be unaffected by all of the aforementioned limitations, allowing it to produce greater results.

Conclusion

This potential cancer therapy magnetic hyperthermia uses magnetic nanoparticles to cause localized heating in cancer cells, eventually eradicating them. Imaging, diagnosis, and treatment planning have all benefited from the application of deep learning, a branch of machine learning. Nevertheless, there has only been a small amount of study on deep learning particularly for magnetic hyperthermia. This is probably because both magnetic hyperthermia and deep learning methods are still in their infancy and are continually being developed. A research publication in 2021 used deep learning to forecast the heating effectiveness of magnetic nanoparticles during magnetic hyperthermia based on those particles’ characteristics. The association between the physicochemical characteristics of several magnetic nanoparticles and their heating effectiveness was examined by the researchers using a deep neural network. The findings suggested that deep learning may be a useful method for creating magnetic nanoparticles for hyperthermia therapy since it was able to predict the heating effectiveness of the magnetic nanoparticles with high accuracy. While there has been some preliminary study on the application of deep learning for magnetic hyperthermia, further research is required to fully realize this technology’s potential.

References

1. Dar MS, Akram KB, Sohail A, Arif F, Zabihi F, Yang S, et al. Heat induction in two-dimensional grapheme-Fe3O4 nanohybrids for magnetic hyperthermia applications with artificial neural network modeling. RSC advances. 2021; 11(35): 21702-21715.

   [Crossref] [Google Scholar] [Pubmed]

2. Ciaburro G, Iannace G. Machine learning-based algorithms to knowledge extraction from time series data: A review. Data. 2021; 6(6): 55.

   [Crossref] [Google Scholar]

3. Ciaburro G. Sound event detection in underground parking garage using convolutional neural network. Big Data Cogn Comp. 2020; 4(3): 20.

   [Crossref] [Google Scholar]

4. Caruana R, Niculescu-Mizil A. An empirical comparison of supervised learning algorithms. Proc 23rd Int Mach Learn 2006: 161-168.

   [Crossref] [Google Scholar]     

5. Celebi ME, Aydin K. Unsupervised learning algorithms. Springer. 2016.
6. Azadeh A, Ghaderi SF, Sohrabkhani S. Forecasting electrical consumption by integration of neural network, time series and ANOVA. Appl Math Comput. 2007; 186(2): 1753-1761.

   [Crossref] [Google Scholar]

7. Miller Jr RG. Beyond ANOVA: Basics of applied statistics. CRC press. 1997.
8. Hill T, O'Connor M, Remus W. Neural network models for time series forecasts. Manage Sci. 1996; 42(7): 1082-1092.

   [Crossref] [Google Scholar]

9. Makridakis S, Andersen A, Carbone R, Fildes R, Hibon M, Lewandowski R, et al. The accuracy of extrapolation (time series) methods: Results of a forecasting competition. J Forecast. 1982; 1(2): 111-153.

   [Crossref] [Google Scholar]

10. Zhang GP. Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing. 2003; 50: 159-175.

    [Crossref] [Google Scholar]

11. Contreras J, Espinola R, Nogales FJ, Conejo AJ. ARIMA models to predict next-day electricity prices. IEEE Trans Power Syst. 2003; 18(3): 1014-1120.

    [Crossref] [Google Scholar]

12. Jain A, Kumar AM. Hybrid neural network models for hydrologic time series forecasting. Appl Soft Comput. 2007; 7(2): 585-592.

    [Crossref] [Google Scholar] [Pubmed]

13. Tseng FM, Yu HC, Tzeng GH. Combining neural network model with seasonal time series ARIMA model. Technol Forecast Soc Change. 2002; 69(1): 71-87.

    [Crossref] [Google Scholar]

14. Chen CF, Chang YH, Chang YW. Seasonal ARIMA forecasting of inbound air travel arrivals to Taiwan. Transportmetrica. 2009; 5(2): 125-140.

    [Crossref] [Google Scholar]

15. Khashei M, Bijari M. An artificial neural network (p, d, q) model for timeseries forecasting. Expert Syst Appl. 2010; 37(1): 479-489.

    [Crossref] [Google Scholar]

16. Chaudhuri TD, Ghosh I. Artificial neural network and time series modeling based approach to forecasting the exchange rate in a multivariate framework. arXiv. 2016.

    [Crossref] [Google Scholar]

17. Aras S, Kocakoç ID. A new model selection strategy in time series forecasting with artificial neural networks: IHTS. Neurocomputing. 2016; 174: 974-987.

    [Crossref] [Google Scholar]

18. Doucoure B, Agbossou K, Cardenas A. Time series prediction using artificial wavelet neural network and multi-resolution analysis: Application to wind speed data. Renew Energy. 2016; 92: 202-211.

    [Crossref] [Google Scholar]

19. Chicea D, Rei SM. A fast artificial neural network approach for dynamic light scattering time series processing. Meas Sci Technol. 2018; 29(10): 105201.

    [Crossref] [Google Scholar]

20. Horák J, Krulicky T. Comparison of exponential time series alignment and time series alignment using artificial neural networks by example of prediction of future development of stock prices of a specific company. EDP Sci. 2019; 61: 01006.

    [Crossref] [Google Scholar]

21. Liu H, Tian HQ, Pan DF, Li YF. Forecasting models for wind speed using wavelet, wavelet packet, time series and artificial neural networks. Appl Energy. 2013; 107: 191-208.

    [Crossref] [Google Scholar]

22. Wang CC, Kang Y, Shen PC, Chang YP, Chung YL. Applications of fault diagnosis in rotating machinery by using time series analysis with neural network. Expert Syst Appl. 2010; 37(2): 1696-1702.

    [Crossref] [Google Scholar]

23. Abiodun OI, Jantan A, Omolara AE, Dada KV, Mohamed NA, Arshad H. State-of-the-art in artificial neural network applications: A survey. Heliyon. 2018; 4(11).

    [Crossref] [Google Scholar] [Pubmed]

24. Da Silva IN, Spatti HD, Flauzino AR, Liboni LH, dos Reis Alves SF, da Silva IN, et al. Artificial neural network architectures and training processes. Springer. 2017.
25. Alanis AY, Arana-Daniel N, Lopez-Franco C. Artificial neural networks for engineering applications. Academic press. 2019.
26. Romero VP, Maffei L, Brambilla G, Ciaburro G. Modelling the soundscape quality of urban waterfronts by artificial neural networks. Appl Acoust. 2016; 111: 121-128.

    [Crossref] [Google Scholar]

27. Khosrow-Pour DB. Advanced methodologies and technologies in artificial intelligence, computer simulation, and human-computer interaction. IGI Global. 2018.
28. Al-Massri R, Al-Astel Y, Ziadia H, Mousa DK, Abu-Naser SS. Classification prediction of SBRCTs cancers using artificial neural network. Int J Acad Eng Res. 2018; 2(11): 1-7.

    [Google Scholar]

29. Wang L, Wang Z, Qu H, Liu S. Optimal forecast combination based on neural networks for time series forecasting. Appl Soft Comput. 2018; 66: 1-7.

    [Crossref] [Google Scholar]

30. Qi M, Zhang GP. An investigation of model selection criteria for neural network time series forecasting. Eur J Oper Res. 2001; 132(3): 666-680.

    [Crossref] [Google Scholar]

31. Tealab A, Hefny H, Badr A. Forecasting of nonlinear time series using ANN. Future Computing Inform J. 2017; 2(1): 39-47.

    [Crossref] [Google Scholar]

32. Goodfellow I, Bengio Y, Courville A. Deep learning. MIT press. 2016.
33. Medeiros MC, Pedreira CE. What are the effects of forecasting linear time series with neural networks? Texto para discussão. 2001.

    [Google Scholar]

34. Kayri M. Predictive abilities of Bayesian regularization and Levenberg-Marquardt algorithms in artificial neural networks: A comparative empirical study on social data. Math Comput Appl. 2016; 21(2): 20.

    [Crossref] [Google Scholar]

35. Murphy KP. Machine learning: A probabilistic perspective. MIT press. 2012.
36. Burden F, Winkler D. Bayesian regularization of neural networks. Artificial neural networks: Methods and applications. 2009: 23-42.
37. Broemeling LD. Bayesian inference for stochastic processes. CRC Press. 2017.
38. Akram KB, ul Hassan SM, Ahmed A, Hamayun MA, Rafique M, Manzoor S. Giant ac magnetoconductivity in rGO-Fe3O4 composites. J Magn Magn Mater. 2020; 499: 166174.

    [Crossref] [Google Scholar]