An explainable Artificial Intelligence software system for predicting diabetes.

Implementing diabetes surveillance systems is paramount to mitigate the risk of incurring substantial medical expenses. Currently, blood glucose is measured by minimally invasive methods, which involve extracting a small blood sample and transmitting it to a blood glucose meter. This method is deemed discomforting for individuals who are undergoing it. The present study introduces an Explainable Artificial Intelligence (XAI) system, which aims to create an intelligible machine capable of explaining expected outcomes and decision models. To this end, we analyze abnormal glucose levels by utilizing Bi-directional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN). In this regard, the glucose levels are acquired through the glucose oxidase (GOD) strips placed over the human body. Later, the signal data is converted to the spectrogram images, classified as low glucose, average glucose, and abnormal glucose levels. The labeled spectrogram images are then used to train the individualized monitoring model. The proposed XAI model to track real-time glucose levels uses the XAI-driven architecture in its feature processing. The model's effectiveness is evaluated by analyzing the performance of the proposed model and several evolutionary metrics used in the confusion matrix. The data revealed in the study demonstrate that the proposed model effectively identifies individuals with elevated glucose levels.

Decoding lower-limb kinematic parameters during pedaling tasks using deep learning approaches and EEG.

Stroke is a neurological condition that usually results in the loss of voluntary control of body movements, making it difficult for individuals to perform activities of daily living (ADLs). Brain-computer interfaces (BCIs) integrated into robotic systems, such as motorized mini exercise bikes (MMEBs), have been demonstrated to be suitable for restoring gait-related functions. However, kinematic estimation of continuous motion in BCI systems based on electroencephalography (EEG) remains a challenge for the scientific community. This study proposes a comparative analysis to evaluate two artificial neural network (ANN)-based decoders to estimate three lower-limb kinematic parameters: x- and y-axis position of the ankle and knee joint angle during pedaling tasks. Long short-term memory (LSTM) was used as a recurrent neural network (RNN), which reached Pearson correlation coefficient (PCC) scores close to 0.58 by reconstructing kinematic parameters from the EEG features on the delta band using a time window of 250 ms. These estimates were evaluated through kinematic variance analysis, where our proposed algorithm showed promising results for identifying pedaling and rest periods, which could increase the usability of classification tasks. Additionally, negative linear correlations were found between pedaling speed and decoder performance, thereby indicating that kinematic parameters between slower speeds may be easier to estimate. The results allow concluding that the use of deep learning (DL)-based methods is feasible for the estimation of lower-limb kinematic parameters during pedaling tasks using EEG signals. This study opens new possibilities for implementing controllers most robust for MMEBs and BCIs based on continuous decoding, which may allow for maximizing the degrees of freedom and personalized rehabilitation.

Multiclass Classification of Visual Electroencephalogram Based on Channel Selection, Minimum Norm Estimation Algorithm, and Deep Network Architectures.

This work addresses the challenge of classifying multiclass visual EEG signals into 40 classes for brain-computer interface applications using deep learning architectures. The visual multiclass classification approach offers BCI applications a significant advantage since it allows the supervision of more than one BCI interaction, considering that each class label supervises a BCI task. However, because of the nonlinearity and nonstationarity of EEG signals, using multiclass classification based on EEG features remains a significant challenge for BCI systems. In the present work, mutual information-based discriminant channel selection and minimum-norm estimate algorithms were implemented to select discriminant channels and enhance the EEG data. Hence, deep EEGNet and convolutional recurrent neural networks were separately implemented to classify the EEG data for image visualization into 40 labels. Using the k-fold cross-validation approach, average classification accuracies of 94.8% and 89.8% were obtained by implementing the aforementioned network architectures. The satisfactory results obtained with this method offer a new implementation opportunity for multitask embedded BCI applications utilizing a reduced number of both channels (<50%) and network parameters (<110 K).

Optimizing RNNs for EMG Signal Classification: A Novel Strategy Using Grey Wolf Optimization.

Accurate classification of electromyographic (EMG) signals is vital in biomedical applications. This study evaluates different architectures of recurrent neural networks for the classification of EMG signals associated with five movements of the right upper extremity. A Butterworth filter was implemented for signal preprocessing, followed by segmentation into 250 ms windows, with an overlap of 190 ms. The resulting dataset was divided into training, validation, and testing subsets. The Grey Wolf Optimization algorithm was applied to the gated recurrent unit (GRU), long short-term memory (LSTM) architectures, and bidirectional recurrent neural networks. In parallel, a performance comparison with support vector machines (SVMs) was performed. The results obtained in the first experimental phase revealed that all the RNN networks evaluated reached a 100% accuracy, standing above the 93% achieved by the SVM. Regarding classification speed, LSTM ranked as the fastest architecture, recording a time of 0.12 ms, followed by GRU with 0.134 ms. Bidirectional recurrent neural networks showed a response time of 0.2 ms, while SVM had the longest time at 2.7 ms. In the second experimental phase, a slight decrease in the accuracy of the RNN models was observed, standing at 98.46% for LSTM, 96.38% for GRU, and 97.63% for the bidirectional network. The findings of this study highlight the effectiveness and speed of recurrent neural networks in the EMG signal classification task.

Spectral temporal graph neural network for multivariate agricultural price forecasting / Rede neural de gráfico temporal espectral para previsão de preços agrícolas multivariados

Multivariate time series forecasting has an important role in many real-world domains. Especially, price prediction has always been on the focus of researchers. Yet, it is a challenging task that requires the capturing of intra-series and inter-series correlations. Most of the models in literature focus only on the correlation in temporal domain. In this paper, we have curated a new dataset from the official website of Turkish Ministry of Commerce. The dataset consists of daily prices and trade volume of vegetables and covers 1791 days between January 1, 2018 and November 26, 2022. A Spectral Temporal Graph Neural Network (StemGNN) is employed on the curated dataset and the results are given in comparison to Convolutional neural networks (CNN), Long short-term memory (LSTM) and Random Forest models. GNN architecture achieved a state-of-the-art result such as mean absolute error (MAE): 1,37 and root mean squared error (RMSE): 1.94). To our knowledge, this is one of the few studies that investigates GNN for time series analysis and the first study in architecture field.

A previsão multivariada de séries temporais tem um papel importante em muitos domínios do mundo real. Especialmente, a previsão de preços sempre esteve no foco dos pesquisadores. No entanto, é uma tarefa desafiadora que requer a captura de correlações intra-séries e inter-séries. A maioria dos modelos na literatura foca apenas a correlação no domínio temporal. Neste artigo, selecionamos um novo conjunto de dados do site oficial do Ministério do Comércio Turco. O conjunto de dados consiste em preços diários e volume comercial de vegetais e abrange 1.791 dias entre 1º de janeiro de 2018 e 26 de novembro de 2022. Uma Rede Neural de Gráfico Temporal Espectral é empregada no conjunto de dados curado e os resultados são fornecidos em comparação com CNN, LSTM e Modelos de Floresta Aleatória. A arquitetura GNN alcançou um resultado de ponta (MAE: 1,37, RMSE: 1,94). Até onde sabemos, este é um dos poucos estudos que investiga GNN para análise de séries temporais e o primeiro estudo na área de arquitetura.

Design of a Soft Sensor Based on Long Short-Term Memory Artificial Neural Network (LSTM) for Wastewater Treatment Plants.

Assessment of wastewater effluent quality in terms of physicochemical and microbial parameters is a difficult task; therefore, an online method which combines the variables and represents a final value as the quality index could be used as a useful management tool for decision makers. However, conventional measurement methods often have limitations, such as time-consuming processes and high associated costs, which hinder efficient and practical monitoring. Therefore, this study presents an approach that underscores the importance of using both short- and long-term memory networks (LSTM) to enhance monitoring capabilities within wastewater treatment plants (WWTPs). The use of LSTM networks for soft sensor design is presented as a promising solution for accurate variable estimation to quantify effluent quality using the total chemical oxygen demand (TCOD) quality index. For the realization of this work, we first generated a dataset that describes the behavior of the activated sludge system in discrete time. Then, we developed a deep LSTM network structure as a basis for formulating the LSTM-based soft sensor model. The results demonstrate that this structure produces high-precision predictions for the concentrations of soluble X1 and solid X2 substrates in the wastewater treatment system. After hyperparameter optimization, the predictive capacity of the proposed model is optimized, with average values of performance metrics, mean square error (MSE), coefficient of determination (R2), and mean absolute percentage error (MAPE), of 23.38, 0.97, and 1.31 for X1, and 9.74, 0.93, and 1.89 for X2, respectively. According to the results, the proposed LSTM-based soft sensor can be a valuable tool for determining effluent quality index in wastewater treatment systems.

The Public Health Contribution of Sentiment Analysis of Monkeypox Tweets to Detect Polarities Using the CNN-LSTM Model.

Monkeypox is a rare disease caused by the monkeypox virus. This disease was considered eradicated in 1980 and was believed to affect rodents and not humans. However, recent years have seen a massive outbreak of monkeypox in humans, setting off worldwide alerts from health agencies. As of September 2022, the number of confirmed cases in Peru had reached 1964. Although most monkeypox patients have been discharged, we cannot neglect the monitoring of the population with respect to the monkeypox virus. Lately, the population has started to express their feelings and opinions through social media, specifically Twitter, as it is the most used social medium and is an ideal space to gather what people think about the monkeypox virus. The information imparted through this medium can be in different formats, such as text, videos, images, audio, etc. The objective of this work is to analyze the positive, negative, and neutral feelings of people who publish their opinions on Twitter with the hashtag #Monkeypox. To find out what people think about this disease, a hybrid-based model architecture built on CNN and LSTM was used to determine the prediction accuracy. The prediction result obtained from the total monkeypox data was 83% accurate. Other performance metrics were also used to evaluate the model, such as specificity, recall level, and F1 score, representing 99%, 85%, and 88%, respectively. The results also showed the polarity of feelings through the CNN-LSTM confusion matrix, where 45.42% of people expressed neither positive nor negative opinions, while 19.45% expressed negative and fearful feelings about this infectious disease. The results of this work contribute to raising public awareness about the monkeypox virus.

Identification of Relevant Protein Interactions with Partial Knowledge: A Complex Network and Deep Learning Approach.

Protein-protein interactions (PPIs) are the basis for understanding most cellular events in biological systems. Several experimental methods, e.g., biochemical, molecular, and genetic methods, have been used to identify protein-protein associations. However, some of them, such as mass spectrometry, are time-consuming and expensive. Machine learning (ML) techniques have been widely used to characterize PPIs, increasing the number of proteins analyzed simultaneously and optimizing time and resources for identifying and predicting protein-protein functional linkages. Previous ML approaches have focused on well-known networks or specific targets but not on identifying relevant proteins with partial or null knowledge of the interaction networks. The proposed approach aims to generate a relevant protein sequence based on bidirectional Long-Short Term Memory (LSTM) with partial knowledge of interactions. The general framework comprises conducting a scale-free and fractal complex network analysis. The outcome of these analyses is then used to fine-tune the fractal method for the vital protein extraction of PPI networks. The results show that several PPI networks are self-similar or fractal, but that both features cannot coexist. The generated protein sequences (by the bidirectional LSTM) also contain an average of 39.5% of proteins in the original sequence. The average length of the generated sequences was 17% of the original one. Finally, 95% of the generated sequences were true.

Deep Learning Multi-Class Approach for Human Fall Detection Based on Doppler Signatures.

Falling events are a global health concern with short- and long-term physical and psychological implications, especially for the elderly population. This work aims to monitor human activity in an indoor environment and recognize falling events without requiring users to carry a device or sensor on their bodies. A sensing platform based on the transmission of a continuous wave (CW) radio-frequency (RF) probe signal was developed using general-purpose equipment. The CW probe signal is similar to the pilot subcarriers transmitted by commercial off-the-shelf WiFi devices. As a result, our methodology can easily be integrated into a joint radio sensing and communication scheme. The sensing process is carried out by analyzing the changes in phase, amplitude, and frequency that the probe signal suffers when it is reflected or scattered by static and moving bodies. These features are commonly extracted from the channel state information (CSI) of WiFi signals. However, CSI relies on complex data acquisition and channel estimation processes. Doppler radars have also been used to monitor human activity. While effective, a radar-based fall detection system requires dedicated hardware. In this paper, we follow an alternative method to characterize falling events on the basis of the Doppler signatures imprinted on the CW probe signal by a falling person. A multi-class deep learning framework for classification was conceived to differentiate falling events from other activities that can be performed in indoor environments. Two neural network models were implemented. The first is based on a long-short-term memory network (LSTM) and the second on a convolutional neural network (CNN). A series of experiments comprising 11 subjects were conducted to collect empirical data and test the system's performance. Falls were detected with an accuracy of 92.1% for the LSTM case, while for the CNN, an accuracy rate of 92.1% was obtained. The results demonstrate the viability of human fall detection based on a radio sensing system such as the one described in this paper.

Application of AI for Short-Term PV Generation Forecast.

The efficient use of the photovoltaic power requires a good estimation of the PV generation. That is why the use of good techniques for forecast is necessary. In this research paper, Long Short-Term Memory, Bidirectional Long Short-Term Memory and the Temporal convolutional network are studied in depth to forecast the photovoltaic power, voltage and efficiency of a 1320 Wp amorphous plant installed in the Technology Support Centre in the University Rey Juan Carlos, Madrid (Spain). The accuracy of these techniques are compared using experimental data along one year, applying 1 timestep or 15 min and 96 step times or 24 h, showing that TCN exhibits outstanding performance, compared with the two other techniques. For instance, it presents better results in all forecast variables and both forecast horizons, achieving an overall Mean Squared Error (MSE) of 0.0024 for 15 min forecasts and 0.0058 for 24 h forecasts. In addition, the sensitivity analyses for the TCN technique is performed and shows that the accuracy is reduced as the forecast horizon increases and that the 6 months of dataset is sufficient to obtain an adequate result with an MSE value of 0.0080 and a coefficient of determination of 0.90 in the worst scenarios (24 h of forecast).

Fault Prediction Based on Leakage Current in Contaminated Insulators Using Enhanced Time Series Forecasting Models.

To improve the monitoring of the electrical power grid, it is necessary to evaluate the influence of contamination in relation to leakage current and its progression to a disruptive discharge. In this paper, insulators were tested in a saline chamber to simulate the increase of salt contamination on their surface. From the time series forecasting of the leakage current, it is possible to evaluate the development of the fault before a flashover occurs. In this paper, for a complete evaluation, the long short-term memory (LSTM), group method of data handling (GMDH), adaptive neuro-fuzzy inference system (ANFIS), bootstrap aggregation (bagging), sequential learning (boosting), random subspace, and stacked generalization (stacking) ensemble learning models are analyzed. From the results of the best structure of the models, the hyperparameters are evaluated and the wavelet transform is used to obtain an enhanced model. The contribution of this paper is related to the improvement of well-established models using the wavelet transform, thus obtaining hybrid models that can be used for several applications. The results showed that using the wavelet transform leads to an improvement in all the used models, especially the wavelet ANFIS model, which had a mean RMSE of 1.58 ×10-3, being the model that had the best result. Furthermore, the results for the standard deviation were 2.18 ×10-19, showing that the model is stable and robust for the application under study. Future work can be performed using other components of the distribution power grid susceptible to contamination because they are installed outdoors.

Attention-Based Deep Recurrent Neural Network to Forecast the Temperature Behavior of an Electric Arc Furnace Side-Wall.

Structural health monitoring (SHM) in an electric arc furnace is performed in several ways. It depends on the kind of element or variable to monitor. For instance, the lining of these furnaces is made of refractory materials that can be worn out over time. Therefore, monitoring the temperatures on the walls and the cooling elements of the furnace is essential for correct structural monitoring. In this work, a multivariate time series temperature prediction was performed through a deep learning approach. To take advantage of data from the last 5 years while not neglecting the initial parts of the sequence in the oldest years, an attention mechanism was used to model time series forecasting using deep learning. The attention mechanism was built on the foundation of the encoder-decoder approach in neural networks. Thus, with the use of an attention mechanism, the long-term dependency of the temperature predictions in a furnace was improved. A warm-up period in the training process of the neural network was implemented. The results of the attention-based mechanism were compared with the use of recurrent neural network architectures to deal with time series data, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The results of the Average Root Mean Square Error (ARMSE) obtained with the attention-based mechanism were the lowest. Finally, a variable importance study was performed to identify the best variables to train the model.

Transfer Learning for COVID-19 cases and deaths forecast using LSTM network.

In this paper, Transfer Learning is used in LSTM networks to forecast new COVID cases and deaths. Models trained in data from early COVID infected countries like Italy and the United States are used to forecast the spread in other countries. Single and multistep forecasting is performed from these models. The results from these models are tested with data from Germany, France, Brazil, India, and Nepal to check the validity of the method. The obtained forecasts are promising and can be helpful for policymakers coping with the threats of COVID-19.

Prediction of malaria using deep learning models: A case study on city clusters in the state of Amazonas, Brazil, from 2003 to 2018

ABSTRACT Background: Malaria is curable. Nonetheless, over 229 million cases of malaria were recorded in 2019, along with 409,000 deaths. Although over 42 million Brazilians are at risk of contracting malaria, 99% percent of all malaria cases in Brazil are located in or around the Amazon rainforest. Despite declining cases and deaths, malaria remains a major public health issue in Brazil. Accurate spatiotemporal prediction of malaria propagation may enable improved resource allocation to support efforts to eradicate the disease. Methods: In response to calls for novel research on malaria elimination strategies that suit local conditions, in this study, we propose machine learning (ML) and deep learning (DL) models to predict the probability of malaria cases in the state of Amazonas. Using a dataset of approximately 6 million records (January 2003 to December 2018), we applied k-means clustering to group cities based on their similarity of malaria incidence. We evaluated random forest, long-short term memory (LSTM) and dated recurrent unit (GRU) models and compared their performance. Results: The LSTM architecture achieved better performance in clusters with less variability in the number of cases, whereas the GRU presents better results in clusters with high variability. Although Diebold-Mariano testing suggested that both the LSTM and GRU performed comparably, GRU can be trained significantly faster, which could prove advantageous in practice. Conclusions: All models showed satisfactory accuracy and strong performance in predicting new cases of malaria, and each could serve as a supplemental tool to support regional policies and strategies.

Tourist Experiences Recommender System Based on Emotion Recognition with Wearable Data.

The collection of physiological data from people has been facilitated due to the mass use of cheap wearable devices. Although the accuracy is low compared to specialized healthcare devices, these can be widely applied in other contexts. This study proposes the architecture for a tourist experiences recommender system (TERS) based on the user's emotional states who wear these devices. The issue lies in detecting emotion from Heart Rate (HR) measurements obtained from these wearables. Unlike most state-of-the-art studies, which have elicited emotions in controlled experiments and with high-accuracy sensors, this research's challenge consisted of emotion recognition (ER) in the daily life context of users based on the gathering of HR data. Furthermore, an objective was to generate the tourist recommendation considering the emotional state of the device wearer. The method used comprises three main phases: The first was the collection of HR measurements and labeling emotions through mobile applications. The second was emotional detection using deep learning algorithms. The final phase was the design and validation of the TERS-ER. In this way, a dataset of HR measurements labeled with emotions was obtained as results. Among the different algorithms tested for ER, the hybrid model of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks had promising results. Moreover, concerning TERS, Collaborative Filtering (CF) using CNN showed better performance.

Assessing the Effect of Climate Variables on the Incidence of Dengue Cases in the Metropolitan Region of Panama City.

The present analysis uses the data of confirmed incidence of dengue cases in the metropolitan region of Panama from 1999 to 2017 and climatic variables (air temperature, precipitation, and relative humidity) during the same period to determine if there exists a correlation between these variables. In addition, we compare the predictive performance of two regression models (SARIMA, SARIMAX) and a recurrent neural network model (RNN-LSTM) on the dengue incidence series. For this data from 1999-2014 was used for training and the three subsequent years of incidence 2015-2017 were used for prediction. The results show a correlation coefficient between the climatic variables and the incidence of dengue were low but statistical significant. The RMSE and MAPE obtained for the SARIMAX and RNN-LSTM models were 25.76, 108.44 and 26.16, 59.68, which suggest that any of these models can be used to predict new outbreaks. Although, it can be said that there is a limited role of climatic variables in the outputs the models. The value of this work is that it helps understand the behaviour of cases in a tropical setting as is the Metropolitan Region of Panama City, and provides the basis needed for a much needed early alert system for the region.

National Holidays and Social Mobility Behaviors: Alternatives for Forecasting COVID-19 Deaths in Brazil.

In this paper, we investigate the influence of holidays and community mobility on the transmission rate and death count of COVID-19 in Brazil. We identify national holidays and hallmark holidays to assess their effect on disease reports of confirmed cases and deaths. First, we use a one-variate model with the number of infected people as input data to forecast the number of deaths. This simple model is compared with a more robust deep learning multi-variate model that uses mobility and transmission rates (R0, Re) from a SEIRD model as input data. A principal components model of community mobility, generated by the principal component analysis (PCA) method, is added to improve the input features for the multi-variate model. The deep learning model architecture is an LSTM stacked layer combined with a dense layer to regress daily deaths caused by COVID-19. The multi-variate model incremented with engineered input features can enhance the forecast performance by up to 18.99% compared to the standard one-variate data-driven model.

A novel deep interval type-2 fuzzy LSTM (DIT2FLSTM) model applied to COVID-19 pandemic time-series prediction.

Currently, the novel COVID-19 coronavirus has been widely spread as a global pandemic. The COVID-19 pandemic has a major influence on human life, healthcare systems, and the economy. There are a large number of methods available for predicting the incidence of the virus. A complex and non-stationary problem such as the COVID-19 pandemic is characterized by high levels of uncertainty in its behavior during the pandemic time. The fuzzy logic, especially Type-2 Fuzzy Logic, is a robust and capable model to cope with high-order uncertainties associated with non-stationary time-dependent features. The main objective of the current study is to present a novel Deep Interval Type-2 Fuzzy LSTM (DIT2FLSTM) model for prediction of the COVID-19 incidence, including new cases, recovery cases, and mortality rate in both short and long time series. The proposed model was evaluated on real datasets produced by the world health organization (WHO) on top highly risked countries, including the USA, Brazil, Russia, India, Peru, Spain, Italy, Iran, Germany, and the U.K. The results confirm the superiority of the DIT2FLSTM model with an average area under the ROC curve (AUC) of 96% and a 95% confidence interval of [92-97] % in the short-term and long-term. The DIT2FLSTM was applied to a well-known standard benchmark, the Mackey-Glass time-series, to show the robustness and proficiency of the proposed model in uncertain and chaotic time series problems. The results were evaluated using a 10-fold cross-validation technique and statistically validated through the t-test method. The proposed DIT2FLSTM model is promising for the prediction of complex problems such as the COVID-19 pandemic and making strategic prevention decisions to save more lives.

Pulmonary COVID-19: Learning Spatiotemporal Features Combining CNN and LSTM Networks for Lung Ultrasound Video Classification.

Deep Learning is a very active and important area for building Computer-Aided Diagnosis (CAD) applications. This work aims to present a hybrid model to classify lung ultrasound (LUS) videos captured by convex transducers to diagnose COVID-19. A Convolutional Neural Network (CNN) performed the extraction of spatial features, and the temporal dependence was learned using a Long Short-Term Memory (LSTM). Different types of convolutional architectures were used for feature extraction. The hybrid model (CNN-LSTM) hyperparameters were optimized using the Optuna framework. The best hybrid model was composed of an Xception pre-trained on ImageNet and an LSTM containing 512 units, configured with a dropout rate of 0.4, two fully connected layers containing 1024 neurons each, and a sequence of 20 frames in the input layer (20×2018). The model presented an average accuracy of 93% and sensitivity of 97% for COVID-19, outperforming models based purely on spatial approaches. Furthermore, feature extraction using transfer learning with models pre-trained on ImageNet provided comparable results to models pre-trained on LUS images. The results corroborate with other studies showing that this model for LUS classification can be an important tool in the fight against COVID-19 and other lung diseases.

Industrial Semi-Supervised Dynamic Soft-Sensor Modeling Approach Based on Deep Relevant Representation Learning.

Soft sensors based on deep learning have been growing in industrial process applications, inferring hard-to-measure but crucial quality-related variables. However, applications may present strong non-linearity, dynamicity, and a lack of labeled data. To deal with the above-cited problems, the extraction of relevant features is becoming a field of interest in soft-sensing. A novel deep representative learning soft-sensor modeling approach is proposed based on stacked autoencoder (SAE), mutual information (MI), and long-short term memory (LSTM). SAE is trained layer by layer with MI evaluation performed between extracted features and targeted output to evaluate the relevance of learned representation in each layer. This approach highlights relevant information and eliminates irrelevant information from the current layer. Thus, deep output-related representative features are retrieved. In the supervised fine-tuning stage, an LSTM is coupled to the tail of the SAE to address system inherent dynamic behavior. Also, a k-fold cross-validation ensemble strategy is applied to enhance the soft-sensor reliability. Two real-world industrial non-linear processes are employed to evaluate the proposed method performance. The obtained results show improved prediction performance in comparison to other traditional and state-of-art methods. Compared to the other methods, the proposed model can generate more than 38.6% and 39.4% improvement of RMSE for the two analyzed industrial cases.