Deep learning models to predict primary open-angle glaucoma.

Glaucoma is a major cause of blindness and vision impairment worldwide, and visual field (VF) tests are essential for monitoring the conversion of glaucoma. While previous studies have primarily focused on using VF data at a single time point for glaucoma prediction, there has been limited exploration of longitudinal trajectories. Additionally, many deep learning techniques treat the time-to-glaucoma prediction as a binary classification problem (glaucoma Yes/No), resulting in the misclassification of some censored subjects into the nonglaucoma category and decreased power. To tackle these challenges, we propose and implement several deep-learning approaches that naturally incorporate temporal and spatial information from longitudinal VF data to predict time-to-glaucoma. When evaluated on the Ocular Hypertension Treatment Study (OHTS) dataset, our proposed convolutional neural network (CNN)-long short-term memory (LSTM) emerged as the top-performing model among all those examined. The implementation code can be found online (https://github.com/rivenzhou/VF_prediction).

EXPRESS: Navigating the Mind's Eye: Understanding Gaze Shifts in Visuospatial Bootstrapping.

Visuo-spatial bootstrapping refers to the well-replicated phenomena in which serial recall in a purely verbal task is boosted by presenting digits within the familiar spatial layout of a typical telephone keypad. The visuo-spatial bootstrapping phenomena indicates that additional support comes from long-term knowledge of a fixed spatial pattern, and prior experimentation supports the idea that access to this benefit depends on the availability of the visuo-spatial motor system (e.g., Allen et al., 2015). We investigate this by tracking participants' eye movements during encoding and retention of verbal lists to learn whether gaze patterns support verbal memory differently when verbal information is presented in the familiar visual layout. Participants' gaze was recorded during attempts to recall lists of seven digits in three formats: centre of the screen, typical telephone keypad, or a spatially identical layout with randomized number placement. Performance was better with the typical than with the novel layout. Our data show that eye movements differ when encoding and retaining verbal information that has a familiar layout compared with the same verbal information presented in a novel layout, suggesting recruitment of different spatial rehearsal strategies. However, no clear link between gaze pattern and recall accuracy was observed, which suggests that gazes play a limited role in retention, at best.

MCMVDRP: a multi-channel multi-view deep learning framework for cancer drug response prediction.

Drug therapy remains the primary approach to treating tumours. Variability among cancer patients, including variations in genomic profiles, often results in divergent therapeutic responses to analogous anti-cancer drug treatments within the same cohort of cancer patients. Hence, predicting the drug response by analysing the genomic profile characteristics of individual patients holds significant research importance. With the notable progress in machine learning and deep learning, many effective methods have emerged for predicting drug responses utilizing features from both drugs and cell lines. However, these methods are inadequate in capturing a sufficient number of features inherent to drugs. Consequently, we propose a representational approach for drugs that incorporates three distinct types of features: the molecular graph, the SMILE strings, and the molecular fingerprints. In this study, a novel deep learning model, named MCMVDRP, is introduced for the prediction of cancer drug responses. In our proposed model, an amalgamation of these extracted features is performed, followed by the utilization of fully connected layers to predict the drug response based on the IC50 values. Experimental results demonstrate that the presented model outperforms current state-of-the-art models in performance.

Promoted production of Fe(IV)/Fe(V) intermediates in the calcium peroxide/ferrate(VI) process for low-damage removal of algal contaminants and membrane fouling control.

Ultrafiltration (UF) is widely employed for harmful algae rejection, whereas severe membrane fouling hampers its long-term operation. Herein, calcium peroxide (CaO2) and ferrate (Fe(VI)) were innovatively coupled for low-damage removal of algal contaminants and fouling control in the UF process. As a result, the terminal J/J0 increased from 0.13 to 0.66, with Rr and Rir respectively decreased by 96.74 % and 48.47 %. The cake layer filtration was significantly postponed, and pore blocking was reduced. The ζ-potential of algal foulants was weakened from -34.4 mV to -18.7 mV, and algal cells of 86.15 % were removed with flocs of 300 µm generated. The cell integrity was better remained in comparison to the Fe(VI) treatment, and Fe(IV)/Fe(V) was verified to be the dominant reactive species. The membrane fouling alleviation mechanisms could be attributed to the reduction of the fouling loads and the changes in the interfacial free energies. A membrane fouling prediction model was built based on a long short-term memory deep learning network, which predicted that the filtration volume at J/J0= 0.2 increased from 288 to 1400 mL. The results provide a new routine for controlling algal membrane fouling from the perspective of promoting the generation of Fe(IV)/Fe(V) intermediates.

ERABiLNet: enhanced residual attention with bidirectional long short-term memory.

Alzheimer's Disease (AD) causes slow death in brain cells due to shrinkage of brain cells which is more prevalent in older people. In most cases, the symptoms of AD are mistaken as age-related stresses. The most widely utilized method to detect AD is Magnetic Resonance Imaging (MRI). Along with Artificial Intelligence (AI) techniques, the efficacy of identifying diseases related to the brain has become easier. But, the identical phenotype makes it challenging to identify the disease from the neuro-images. Hence, a deep learning method to detect AD at the beginning stage is suggested in this work. The newly implemented "Enhanced Residual Attention with Bi-directional Long Short-Term Memory (Bi-LSTM) (ERABi-LNet)" is used in the detection phase to identify the AD from the MRI images. This model is used for enhancing the performance of the Alzheimer's detection in scale of 2-5%, minimizing the error rates, increasing the balance of the model, so that the multi-class problems are supported. At first, MRI images are given to "Residual Attention Network (RAN)", which is specially developed with three convolutional layers, namely atrous, dilated and Depth-Wise Separable (DWS), to obtain the relevant attributes. The most appropriate attributes are determined by these layers, and subjected to target-based fusion. Then the fused attributes are fed into the "Attention-based Bi-LSTM". The final outcome is obtained from this unit. The detection efficiency based on median is 26.37% and accuracy is 97.367% obtained by tuning the parameters in the ERABi-LNet with the help of Modified Search and Rescue Operations (MCDMR-SRO). The obtained results are compared with ROA-ERABi-LNet, EOO-ERABi-LNet, GTBO-ERABi-LNet and SRO-ERABi-LNet respectively. The ERABi_LNet thus provides enhanced accuracy and other performance metrics compared to such deep learning models. The proposed method has the better sensitivity, specificity, F1-Score and False Positive Rate compared with all the above mentioned competing models with values such as 97.49%.97.84%,97.74% and 2.616 respective;y. This ensures that the model has better learning capabilities and provides lesser false positives with balanced prediction.

Hybrid deep learning based prediction for water quality of plain watershed.

Establishing a highly reliable and accurate water quality prediction model is critical for effective water environment management. However, enhancing the performance of these predictive models continues to pose challenges, especially in the plain watershed with complex hydraulic conditions. This study aims to evaluate the efficacy of three traditional machine learning models versus three deep learning models in predicting the water quality of plain river networks and to develop a novel hybrid deep learning model to further improve prediction accuracy. The performance of the proposed model was assessed under various input feature sets and data temporal frequencies. The findings indicated that deep learning models outperformed traditional machine learning models in handling complex time series data. Long Short-Term Memory (LSTM) models improved the R2 by approximately 29% and lowered the Root Mean Square Error (RMSE) by about 48.6% on average. The hybrid Bayes-LSTM-GRU (Gated Recurrent Unit) model significantly enhanced prediction accuracy, reducing the average RMSE by 18.1% compared to the single LSTM model. Models trained on feature-selected datasets exhibited superior performance compared to those trained on original datasets. Higher temporal frequencies of input data generally provide more useful information. However, in datasets with numerous abrupt changes, increasing the temporal interval proves beneficial. Overall, the proposed hybrid deep learning model demonstrates an efficient and cost-effective method for improving water quality prediction performance, showing significant potential for application in managing water quality in plain watershed.

Exploring the contribution of joint angles and sEMG signals on joint torque prediction accuracy using LSTM-based deep learning techniques.

Machine learning (ML) has been used to predict lower extremity joint torques from joint angles and surface electromyography (sEMG) signals. This study trained three bidirectional Long Short-Term Memory (LSTM) models, which utilize joint angle, sEMG, and combined modalities as inputs, using a publicly accessible dataset to estimate joint torques during normal walking and assessed the performance of models, that used specific inputs independently plus the accuracy of the joint-specific torque prediction. The performance of each model was evaluated using normalized root mean square error (nRMSE) and Pearson correlation coefficient (PCC). Each model's median scores for the PCC and nRMSE values were highly convergent and the bulk of the mean nRMSE values of all joints were less than 10%. The ankle joint torque was the most successfully predicted output, having a mean nRMSE of less than 9% for all models. The knee joint torque prediction has reached the highest accuracy with a mean nRMSE of 11% and the hip joint torque prediction of 10%. The PCC values of each model were significantly high and remarkably comparable for the ankle (∼ 0.98), knee (∼ 0.92), and hip (∼ 0.95) joints. The model obtained significantly close accuracy with single and combined input modalities, indicating that one of either input may be sufficient for predicting the torque of a particular joint, obviating the need for the other in certain contexts.

Influence of presentation duration on filtering of irrelevant stimuli in visual working memory.

In environments teeming with distractions, the ability to selectively focus on relevant information is crucial for advanced cognitive processing. Existing research using event-related potential (ERP) technology has shown active suppression of irrelevant stimuli during the consolidation phase of visual working memory (VWM). In previous studies, participants have always been given sufficient time to consolidate VWM, while suppressing distracting information. However, it remains unclear whether the suppression of irrelevant distractors requires continuous effort throughout their presence or whether this suppression is only necessary after the consolidation of task-relevant information. To address this question, our study examines whether distractor suppression is necessary in scenarios where consolidation time is limited. This research investigates the effect of varying presentation durations on the filtering of distractors in VWM. We tasked participants with memorizing two color stimuli and ignoring four distractors, presented for either 50 ms or 200 ms. Using ERP technology, we discovered that the distractor-induced distractor positivity (PD) amplitude is larger during longer presentation durations compared to shorter ones. These findings underscore the significant impact of presentation duration on the efficacy of distractor suppression in VWM, as prolonged exposure results in a stronger suppression effect on distractors. This study sheds light on the temporal dynamics of attention and memory, emphasizing the critical role of stimulus timing in cognitive tasks. These findings provide valuable insights into the mechanisms underlying VWM and have significant implications for models of attention and memory.

CNN-BLSTM based deep learning framework for eukaryotic kinome classification: An explainability based approach.

Classification of protein families from their sequences is an enduring task in Proteomics and related studies. Numerous deep-learning models have been moulded to tackle this challenge, but due to the black-box character, they still fall short in reliability. Here, we present a novel explainability pipeline that explains the pivotal decisions of the deep learning model on the classification of the Eukaryotic kinome. Based on a comparative and experimental analysis of the most cutting-edge deep learning algorithms, the best deep learning model CNN-BLSTM was chosen to classify the eight eukaryotic kinase sequences to their corresponding families. As a substitution for the conventional class activation map-based interpretation of CNN-based models in the domain, we have cascaded the GRAD CAM and Integrated Gradient (IG) explainability modus operandi for improved and responsible results. To ensure the trustworthiness of the classifier, we have masked the kinase domain traces, identified from the explainability pipeline and observed a class-specific drop in F1-score from 0.96 to 0.76. In compliance with the Explainable AI paradigm, our results are promising and contribute to enhancing the trustworthiness of deep learning models for biological sequence-associated studies.

GCN-Based LSTM Autoencoder with Self-Attention for Bearing Fault Diagnosis.

The manufacturing industry has been operating within a constantly evolving technological environment, underscoring the importance of maintaining the efficiency and reliability of manufacturing processes. Motor-related failures, especially bearing defects, are common and serious issues in manufacturing processes. Bearings provide accurate and smooth movements and play essential roles in mechanical equipment with shafts. Given their importance, bearing failure diagnosis has been extensively studied. However, the imbalance in failure data and the complexity of time series data make diagnosis challenging. Conventional AI models (convolutional neural networks (CNNs), long short-term memory (LSTM), support vector machine (SVM), and extreme gradient boosting (XGBoost)) face limitations in diagnosing such failures. To address this problem, this paper proposes a bearing failure diagnosis model using a graph convolution network (GCN)-based LSTM autoencoder with self-attention. The model was trained on data extracted from the Case Western Reserve University (CWRU) dataset and a fault simulator testbed. The proposed model achieved 97.3% accuracy on the CWRU dataset and 99.9% accuracy on the fault simulator dataset.

The Prediction and Evaluation of Surface Quality during the Milling of Blade-Root Grooves Based on a Long Short-Term Memory Network and Signal Fusion.

The surface quality of milled blade-root grooves in industrial turbine blades significantly influences their mechanical properties. The surface texture reveals the interaction between the tool and the workpiece during the machining process, which plays a key role in determining the surface quality. In addition, there is a significant correlation between acoustic vibration signals and surface texture features. However, current research on surface quality is still relatively limited, and most considers only a single signal. In this paper, 160 sets of industrial field data were collected by multiple sensors to study the surface quality of a blade-root groove. A surface texture feature prediction method based on acoustic vibration signal fusion is proposed to evaluate the surface quality. Fast Fourier transform (FFT) is used to process the signal, and the clean and smooth features are extracted by combining wavelet denoising and multivariate smoothing denoising. At the same time, based on the gray-level co-occurrence matrix, the surface texture image features of different angles of the blade-root groove are extracted to describe the texture features. The fused acoustic vibration signal features are input, and the texture features are output to establish a texture feature prediction model. After predicting the texture features, the surface quality is evaluated by setting a threshold value. The threshold is selected based on all sample data, and the final judgment accuracy is 90%.

Comparative Analysis of Macular and Optic Disc Perfusion Pre and Post Silicone Oil Removal: A Machine Learning Approach.

In the realm of ophthalmic surgeries, silicone oil is often utilized as a tamponade agent for repairing retinal detachments, but it necessitates subsequent removal. This study harnesses the power of machine learning to analyze the macular and optic disc perfusion changes pre and post-silicone oil removal, using Optical Coherence Tomography Angiography (OCTA) data. Building upon the foundational work of prior research, our investigation employs Gaussian Process Regression (GPR) and Long Short-Term Memory (LSTM) networks to create predictive models based on OCTA scans. We conducted a comparative analysis focusing on the flow in the outer retina and vessel density in the deep capillary plexus (superior-hemi and perifovea) to track perfusion changes across different time points. Our findings indicate that while machine learning models predict the flow in the outer retina with reasonable accuracy, predicting the vessel density in the deep capillary plexus (particularly in the superior-hemi and perifovea regions) remains challenging. These results underscore the potential of machine learning to contribute to personalized patient care in ophthalmology, despite the inherent complexities in predicting ocular perfusion changes.

Efficient state of charge estimation of lithium-ion batteries in electric vehicles using evolutionary intelligence-assisted GLA-CNN-Bi-LSTM deep learning model.

The battery's performance heavily influences the safety, dependability, and operational efficiency of electric vehicles (EVs). This paper introduces an innovative hybrid deep learning architecture that dramatically enhances the estimation of the state of charge (SoC) of lithium-ion (Li-ion) batteries, crucial for efficient EV operation. Our model uniquely integrates a convolutional neural network (CNN) with bidirectional long short-term memory (Bi-LSTM), optimized through evolutionary intelligence, enabling an advanced level of precision in SoC estimation. A novel aspect of this work is the application of the Group Learning Algorithm (GLA) to tune the hyperparameters of the CNN-Bi-LSTM network meticulously. This approach not only refines the model's accuracy but also significantly enhances its efficiency by optimizing each parameter to best capture and integrate both spatial and temporal information from the battery data. This is in stark contrast to conventional models that typically focus on either spatial or temporal data, but not both effectively. The model's robustness is further demonstrated through its training across six diverse datasets that represent a range of EV discharge profiles, including the Highway Fuel Economy Test (HWFET), the US06 test, the Beijing Dynamic Stress Test (BJDST), the dynamic stress test (DST), the federal urban driving schedule (FUDS), and the urban development driving schedule (UDDS). These tests are crucial for ensuring that the model can perform under various real-world conditions. Experimentally, our hybrid model not only surpasses the performance of existing LSTM and CNN frameworks in tracking SoC estimation but also achieves an impressively quick convergence to true SoC values, maintaining an average root mean square error (RMSE) of less than 1 %. Furthermore, the experimental outcomes suggest that this new deep learning methodology outstrips conventional approaches in both convergence speed and estimation accuracy, thus promising to significantly enhance battery life and overall EV efficiency.

Classification of mental workload with EEG analysis by using effective connectivity and a hybrid model of CNN and LSTM.

Estimation of mental workload from electroencephalogram (EEG) signals aims to accurately measure the cognitive demands placed on an individual during multitasking mental activities. By analyzing the brain activity of the subject, we can determine the level of mental effort required to perform a task and optimize the workload to prevent cognitive overload or underload. This information can be used to enhance performance and productivity in various fields such as healthcare, education, and aviation. In this paper, we propose a method that uses EEG and deep neural networks to estimate the mental workload of human subjects during multitasking mental activities. Notably, our proposed method employs subject-independent classification. We use the "STEW" dataset, which consists of two tasks, namely "No task" and "simultaneous capacity (SIMKAP)-based multitasking activity". We estimate the different workload levels of two tasks using a composite framework consisting of brain connectivity and deep neural networks. After the initial preprocessing of EEG signals, an analysis of the relationships between the 14 EEG channels is conducted to evaluate effective brain connectivity. This assessment illustrates the information flow between various brain regions, utilizing the direct Directed Transfer Function (dDTF) method. Then, we propose a deep hybrid model based on pre-trained Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for the classification of workload levels. The accuracy of the proposed deep model achieved 83.12% according to the subject-independent leave-subject-out (LSO) approach. The pre-trained CNN + LSTM approaches to EEG data have been found to be an accurate method for assessing the mental workload.

Cognitive workload estimation using physiological measures: a review.

Estimating cognitive workload levels is an emerging research topic in the cognitive neuroscience domain, as participants' performance is highly influenced by cognitive overload or underload results. Different physiological measures such as Electroencephalography (EEG), Functional Magnetic Resonance Imaging, Functional near-infrared spectroscopy, respiratory activity, and eye activity are efficiently used to estimate workload levels with the help of machine learning or deep learning techniques. Some reviews focus only on EEG-based workload estimation using machine learning classifiers or multimodal fusion of different physiological measures for workload estimation. However, a detailed analysis of all physiological measures for estimating cognitive workload levels still needs to be discovered. Thus, this survey highlights the in-depth analysis of all the physiological measures for assessing cognitive workload. This survey emphasizes the basics of cognitive workload, open-access datasets, the experimental paradigm of cognitive tasks, and different measures for estimating workload levels. Lastly, we emphasize the significant findings from this review and identify the open challenges. In addition, we also specify future scopes for researchers to overcome those challenges.

Independent replications reveal anterior and posterior cingulate cortex activation underlying state anxiety-attenuated face encoding.

Anxiety involves the anticipation of aversive outcomes and can impair neurocognitive processes, such as the ability to recall faces encoded during the anxious state. It is important to precisely delineate and determine the replicability of these effects using causal state anxiety inductions in the general population. This study therefore aimed to replicate prior research on the distinct impacts of threat-of-shock-induced anxiety on the encoding and recognition stage of emotional face processing, in a large asymptomatic sample (n = 92). We successfully replicated previous results demonstrating impaired recognition of faces encoded under threat-of-shock. This was supported by a mega-analysis across three independent studies using the same paradigm (n = 211). Underlying this, a whole-brain fMRI analysis revealed enhanced activation in the posterior cingulate cortex (PCC), alongside previously seen activity in the anterior cingulate cortex (ACC) when combined in a mega-analysis with the fMRI findings we aimed to replicate. We further found replications of hippocampus activation when the retrieval and encoding states were congruent. Our results support the notion that state anxiety disrupts face recognition, potentially due to attentional demands of anxious arousal competing with affective stimuli processing during encoding and suggest that regions of the cingulate cortex play pivotal roles in this.

A novel deep learning identifier for promoters and their strength using heterogeneous features.

Promoters, which are short (50-1500 base-pair) in DNA regions, have emerged to play a critical role in the regulation of gene transcription. Numerous dangerous diseases, likewise cancer, cardiovascular, and inflammatory bowel diseases, are caused by genetic variations in promoters. Consequently, the correct identification and characterization of promoters are significant for the discovery of drugs. However, experimental approaches to recognizing promoters and their strengths are challenging in terms of cost, time, and resources. Therefore, computational techniques are highly desirable for the correct characterization of promoters from unannotated genomic data. Here, we designed a powerful bi-layer deep-learning based predictor named "PROCABLES", which discriminates DNA samples as promoters in the first-phase and strong or weak promoters in the second-phase respectively. The proposed method utilizes five distinct features, such as word2vec, k-spaced nucleotide pairs, trinucleotide propensity-based features, trinucleotide composition, and electron-ion interaction pseudopotentials, to extract the hidden patterns from the DNA sequence. Afterwards, a stacked framework is formed by integrating a convolutional neural network (CNN) with bidirectional long-short-term memory (LSTM) using multi-view attributes to train the proposed model. The PROCABLES model achieved an accuracy of 0.971 and 0.920 and the MCC 0.940 and 0.840 for the first and second-layer using the ten-fold cross-validation test, respectively. The predicted results anticipate that the proposed PROCABLES protocol outperformed the advanced computational predictors targeting promoters and their types. In summary, this research will provide useful hints for the recognition of large-scale promoters in particular and other DNA problems in general.

Independent Effects of Age, Education, Verbal Working Memory, Motor Speed of Processing, Locality, and Morphosyntactic Category on Verb-Related Morphosyntactic Production: Evidence From Healthy Aging.

This study investigates the role of locality (a task/material-related variable), demographic factors (age, education, and sex), cognitive capacities (verbal working memory [WM], verbal short-term memory [STM], speed of processing [SOP], and inhibition), and morphosyntactic category (time reference and grammatical aspect) in verb-related morphosyntactic production (VRMP). A sentence completion task tapping production of time reference and grammatical aspect in local and nonlocal configurations, and cognitive tasks measuring verbal WM capacity, verbal STM capacity, motor SOP, perceptual SOP, and inhibition were administered to 200 neurotypical Greek-speaking participants, aged between 19 and 80 years. We fitted generalized linear mixed-effects models and performed path analyses. Significant main effects of locality, age, education, verbal WM capacity, motor SOP, and morphosyntactic category emerged. Production of time reference and aspect did not interact with any of the significant factors (i.e., age, education, verbal WM capacity, motor SOP, and locality), and locality did not interact with any memory system. Path analyses revealed that the relationships between age and VRMP, and between education and VRMP were partly mediated by verbal WM; and the relationship between verbal WM and VRMP was partly mediated by perceptual SOP. Results suggest that subject-, task/material- and morphosyntactic category-specific factors determine accuracy performance on VRMP; and the effects of age, education, and verbal WM on VRMP are partly indirect. The fact that there was a significant main effect of verbal WM but not of verbal STM on accuracy performance in the VRMP task suggests that it is predominantly the processing component (and not the storage component) of verbal WM that supports VRMP. Lastly, we interpret the results as suggesting that VRMP is also supported by a procedural memory system whose efficiency might be reflected in years of formal education.

Backlight and dim space object detection based on a novel event camera.

Background: For space object detection tasks, conventional optical cameras face various application challenges, including backlight issues and dim light conditions. As a novel optical camera, the event camera has the advantages of high temporal resolution and high dynamic range due to asynchronous output characteristics, which provides a new solution to the above challenges. However, the asynchronous output characteristic of event cameras makes them incompatible with conventional object detection methods designed for frame images. Methods: Asynchronous convolutional memory network (ACMNet) for processing event camera data is proposed to solve the problem of backlight and dim space object detection. The key idea of ACMNet is to first characterize the asynchronous event streams with the Event Spike Tensor (EST) voxel grid through the exponential kernel function, then extract spatial features using a feed-forward feature extraction network, and aggregate temporal features using a proposed convolutional spatiotemporal memory module ConvLSTM, and finally, the end-to-end object detection using continuous event streams is realized. Results: Comparison experiments among ACMNet and classical object detection methods are carried out on Event_DVS_space7, which is a large-scale space synthetic event dataset based on event cameras. The results show that the performance of ACMNet is superior to the others, and the mAP is improved by 12.7% while maintaining the processing speed. Moreover, event cameras still have a good performance in backlight and dim light conditions where conventional optical cameras fail. This research offers a novel possibility for detection under intricate lighting and motion conditions, emphasizing the superior benefits of event cameras in the realm of space object detection.

Recognition of inscribed cursive Pashtu numeral through optimized deep learning.

Pashtu is one of the most widely spoken languages in south-east Asia. Pashtu Numerics recognition poses challenges due to its cursive nature. Despite this, employing a machine learning-based optical character recognition (OCR) model can be an effective way to tackle this issue. The main aim of the study is to propose an optimized machine learning model which can efficiently identify Pashtu numerics from 0-9. The methodology includes data organizing into different directories each representing labels. After that, the data is preprocessed i.e., images are resized to 32 × 32 images, then they are normalized by dividing their pixel value by 255, and the data is reshaped for model input. The dataset was split in the ratio of 80:20. After this, optimized hyperparameters were selected for LSTM and CNN models with the help of trial-and-error technique. Models were evaluated by accuracy and loss graphs, classification report, and confusion matrix. The results indicate that the proposed LSTM model slightly outperforms the proposed CNN model with a macro-average of precision: 0.9877, recall: 0.9876, F1 score: 0.9876. Both models demonstrate remarkable performance in accurately recognizing Pashtu numerics, achieving an accuracy level of nearly 98%. Notably, the LSTM model exhibits a marginal advantage over the CNN model in this regard.