Artificial intelligence (AI), particularly modern deep learning (DL), has transformed an extraordinary range of application domains over the past decade. Breakthroughs in computer vision, natural-language processing (NLP), speech recognition, and reinforcement learning have moved the performance of automated systems beyond human-level on benchmark tasks such as ImageNet image classification (LeCun et al., 2015), conversational language modeling (Brown et al., 2020), and board-game playing (Silver et al., 2018). In health and life sciences, deep neural networks already match or surpass specialist clinicians in skin cancer recognition (Esteva et al., 2017), chest-X-ray triage (Rajpurkar et al., 2017), and diabetic retinopathy screening (Gulshan et al., 2016). These successes illustrate the general ability of AI methods to learn complex, high-dimensional patterns from raw or minimally processed data, providing objective, reproducible, and scalable decision support in settings where conventional analyses remain labor-intensive or inconsistent (Topol, 2019).

A rapidly growing subfield applies AI to neuro-cognitive assessment. Automated analysis of speech, handwriting, and physiological signals has begun to aid the early detection of neurodevelopmental and neurodegenerative disorders such as autism spectrum disorder, hyperactivity disorder, Parkinson's disease, and Alzheimer's disease (Probol and Mieskes, 2024; Nafisah et al., 2025; Gallo-Aristizabal et al., 2025). Within this field, reading disabilities, and dyslexia in particular, represent a compelling target for AI-based diagnostics.

Dyslexia is a specific learning disability of neurobiological origin that impairs reading and language processing skills. It affects an estimated 5%–10% of the population worldwide ( 7% of children, accounting for up to 80% of learning disorders). It is characterized by persistent difficulties in word recognition, decoding, and reading fluency despite adequate intelligence and education (Tiwari et al., 2025; Toki, 2024). Early identification is critical: timely evidence-based intervention can exploit neural plasticity to improve literacy outcomes and mitigate long-term academic and socio-emotional consequences (El Hmimdi et al., 2024). However, conventional diagnostic procedures rely on standardized reading batteries and clinician observation—methods that are time-consuming, subjective, and often unavailable in many educational settings, leading to delayed or inaccurate diagnoses (Snowling, 2013).

Eye-tracking provides a non-invasive, real-time record of visual attention during reading, capturing spatiotemporal gaze trajectories at millisecond resolution. Dyslexic readers consistently produce more frequent and longer fixations, shorter saccades, and an increased number of regressions relative to typically developing peers (Yin et al., 2025). Because these metrics directly reflect underlying cognitive processing demands, eye-movement data furnish a rich, quantitative substrate for automated dyslexia detection.

Machine-learning (ML) approaches leveraging handcrafted eye-movement features have achieved promising performance. Jothi Prabha and Bhargavi (2022) and El Hmimdi et al. (2024) extracted 168 temporal, spatial, and frequency descriptors, applied recursive feature elimination (RFE), and trained a support-vector machine (SVM) that reached 95.6% accuracy. Comparable results have been reported with principal-component analysis and SVM (Nerušil et al., 2021), evolutionary optimization plus SVM (Jothi Prabha and Bhargavi, 2022), k-nearest neighbors and random forests, provided that irrelevant or redundant variables are pruned (Prabha and Bhargavi, 2020). Nonetheless, performance varies with dataset size and reading paradigm, occasionally dropping to ~80% accuracy on smaller or more heterogeneous samples, underscoring challenges of generalization.

DL methods aim to circumvent manual feature engineering by learning hierarchical representations directly from raw gaze sequences. Nerušil et al. (2021) proposed a convolutional neural network (CNN) that ingests (x, y) time-series and implicitly captures fixation–saccade dynamics. The recently released ETDD70 dataset—70 Czech children, 35 with dyslexia—enabled (Sedmidubsky et al., 2024) to obtain ~90% accuracy with a CNN, while Gomolka et al. (2024) achieved 97.7% using a long short-term memory (LSTM) network. Transformer architectures with multi-head attention have pushed reported accuracies even higher (~99%) in tightly controlled experiments (Priyasri and Devi, 2025). Conversely, when trained on a large and diverse corpus of >4, 000 reading traces, a CNN delivered more modest but still encouraging precision and recall of 77%–80% (El Hmimdi et al., 2024), highlighting persistent risks of overfitting and data-shift.

With the advantages achieved by the previous studies, they still have limitations. This motivated us to develop an alternative AI technique that improves the performance of particle swarm optimization (PSO) (Kennedy and Eberhart, 1995). This modification was conducted using the optical computing mechanism (OCM) (Zuo et al., 2019; Lu and Saleh, 1990). In general, OCM depends on photonics that improve and accelerate the computation of complex matrix-vector multiplication. Therefore, the OCM has advantages over electronics, such as high computation speed and parallelism (Kitayama et al., 2019). The increasing complexity of data manipulation techniques and the size of data lead to an increase in demand for highly integrated, scalable optical hardware that is ultracompact and consumes less energy. Therefore, the advantages of ultracompact size lead to the establishment of large, compact computing units. This is considered the main component in optical-artificial-intelligence computers (Feldmann et al., 2021; Kues et al., 2017). This inspired us to simulate the operation of an optical computing mechanism and integrate it with a metaheuristic (MH) technique named PSO. The primary reason for using PSO rather than other MH methods is its broad applicability across diverse applications and its role as the basis for many other MH techniques.

The developed model, named optical PSO (OPSO), begins by generating a set of solutions. Then, it computes the fitness value and determines the best personal and global best values. Next, we simulated the processing of particle-state information through an optical device by encoding the input and performing complex mixing and nonlinearity. The encoding of the input is formulated by computing the optical of the input vector (i.e., current particle, best personal, and global best particle) based on the magnitude and 2π-scaled. The input signal is applied to a coherent linear system using the complex weight matrix, thereby emulating the physical diffraction and interference. Next, a nonlinear saturation of light in a nonlinear crystal is simulated. Thereafter, the optical output is referred to as the optical delta, which is computed as the real part of the nonlinear saturation. Then, the adaptively mixed delta is computed, which depends on the combination of the optical and classical PSO terms using a decaying mixing ratio. Finally, the particle's velocity is updated using the adaptively mixed delta, and a new particle is generated based on the updated velocity. First, the stop conditions are checked to determine whether they are met; then the best solution is used as the output. To assess the performance of the developed model, we applied it to the CEC2019 benchmark functions and used it as a feature selection method to enhance the detection of Dyslexia.

The main contributions of this study can be summarized as follows:

The rest of this study is organized as follows: In Section 2, related work of PSO variants is introduced. In Section 3, the background of traditional Particle Swarm Optimization (PSO) is presented. Section 4 presents the steps of the developed Optical OPSO. Section 5 introduces the experimental results of the proposed model to handle global optimization problems. Whereas Section 6 shows the results of the proposed OPSO to detect dyslexia. The conclusion and future works are given in Section 7.

A recent study on PSO introduced a wide range of variants that adjust the learning strategy, swarm topology, and hybridization mechanisms to improve the balance between exploration and exploitation, increase robustness, and better adapt to specific problem types (Chauhan et al., 2025). Several PSO variants have been developed to automatically adjust the inertia weight, learning factors, or velocity limits to control the search process (Shami et al., 2022), providing parameter control and the bare bones of PSO. ECLPSO and ACLPSO, for example, use adaptive learning probabilities, perturbation-based exploitation, and dimension-wise parameter adaptation to enhance both exploration and convergence compared with the original CLPSO (Lin et al., 2019; Yu and Qiao, 2021). A bare-bones PSO (BBPSO) remains an important minimalist variant that removes velocities and samples positions from probability distributions. It has been extended with deep memory and scale-matrix adaptation to achieve higher precision while maintaining diversity (Song et al., 2021; Sun et al., 2023).

For Learning-Strategy PSO (CLPSO Family; Lin et al., 2019; Liang et al., 2006), one powerful variation supports each particle to learn from the personal-best positions of several peers rather than depending on a single global or local best. Comprehensive learning PSO (CLPSO) enables every particle to learn. This broader information sharing supports diversity and lowers early convergence. Following this concept, ECLPSO, ACLPSO, and CLPSO-OPR introduce adaptive learning probabilities, perturbation operators, and recombination rules in subsequent iterations to more effectively balance broad-area exploration with focused local search (Liang et al., 2006). Recent polls emphasize these learning-oriented PSOs as a significant current trend because they alter how information is shared within the swarm rather than merely tweaking parameter settings (Liang et al., 2006; Chen et al., 2024).

In multiswarm and cooperative PSO, dynamic multi-swarm PSO (DMS-PSO) divides the population into several small sub-swarms that are often regrouped, giving each sub-swarm local autonomy while allowing periodic information exchange to avoid local optima (Zhao et al., 2008). Later research combined DMS-PSO with local search or harmony search to enhance local accuracy and large-scale performance, especially on multimodal benchmarks (Zhao et al., 2011). Many surveys highlight these multi-swarm and collaborative systems as successful in difficult terrain, as they preserve diversity without completely sacrificing convergence speed.

Increasingly, hybrid PSO studies combine BBPSO or classic PSO with problem-specific techniques, including local search, adaptive mutation, and mutual-information-based assessment (Grazioso et al., 2025). For example, a basic PSO with mutual-information-guided feature selection introduces adaptive flip mutation and leader-update rules, enabling efficient dimensionality reduction and competitive classification performance on benchmark datasets. Deep memory BBPSO (DMBBPSO) introduces a unique memory topology and pruning method that allows particles to leverage extensive historical information while eliminating obsolete data, thereby supporting both extensive exploration and high-precision local search (Grazioso et al., 2025). These PSO variants are summarized in Table 1.

The particle swarm optimization (PSO) algorithm is considered one of the most popular and efficient metaheuristic techniques. In general, PSO simulates the behavior of particles in nature during the food search. This includes the communication between populations and how they explore the search space.

To achieve the process of finding the food, the position and velocity of each particle are updated. The first step in PSO is to generate the initial value for the velocity and position of a set of N particles. Then, the fitness value for each particle is computed, and the best personal particle (X_pi) and the global best particle (X_g) are determined. The next process is to update the velocity of each particle (v_i) using Equation 1.

Where c₁ and c₂ refer to the weight coefficient of the best local and global positions, respectively. w is the inertia coefficient, which controls the influence of the previous velocity on the updated velocity. The next step after updating the velocity is updating the particle's position. X_i as in Equation 2.

The process of updating the position is repeated until the stop conditions are reached. Algorithm 1 shows the steps of the PSO algorithm.

The proposed optical Particle Swarm Optimization (OPSO) algorithm is introduced in this section. The proposed OPSO embeds the mathematical properties of all-optical mechanisms within the standard PSO algorithm. This mechanism aims to leverage the inherent properties of coherent optical systems—specifically complex-domain computation and nonlinear light-matter interactions to enhance the search process for particles.

This section presents an ablation study to evaluate the individual contributions of the components within the OPSO algorithm by comparing its performance against the standard PSO on the CEC2019 benchmark functions. The global experimental settings were configured to 1,000 iterations with a population size of 100. The comparative analysis reveals clear advantages of the proposed OPSO method in optimization accuracy as listed in Table 2.

Across all benchmark functions, especially in F1 and F2, the OPSO algorithm performed, on average, better than classical PSO in terms of fitness. This suggests that, due to the properties of CEC2019 functions, OPSO more effectively explores the search space and converges more consistently toward diverse optima across different tests.

To examine the stability of the solutions, the standard deviation merit was used. OPSO across numerous functions yielded less variability in fitness results than PSO, indicating that the runs belonged to a consistent type of stability. This is an important factor to consider in practical applications, where the algorithm's performance must remain stable. This also indicated that the enhancements of the proposed method mitigate the random stochastic disturbances commonly encountered by traditional PSO.

The analysis of the minimum and maximum fitness values confirmed that OPSO kept the ranges between the best and worst results narrower for most functions. With PSO, there was greater variation, especially on problems where the function to be minimized was complex and the search space was high-dimensional. This supports the conclusion that OPSO strikes a better balance between exploration and exploitation and maintains the swarm's diversity, thereby reducing the occurrence of suboptimal solutions in difficult environments.

In terms of computational efficiency, OPSO exhibited performance similar to that of PSO. Although the optimization ability of OPSO is enhanced, this does not entail a significant additional computational burden. Thus, OPSO is suitable for large-scale problems and for applications where time is of the essence and efficiency and accuracy are equally important.

The convergence curves, as presented in Figure 1, indicate that OPSO reaches lower fitness values faster than PSO in most cases, especially on complicated test functions.

To summarize, these results demonstrate the effectiveness of OPSO as a warm-intelligence algorithm. The proposed method showed good accuracy, greater solution stability, and consistent convergence across the CEC2019 benchmarks.

In this study, we presented a modified optimization model by developing the Optical Particle Swarm Optimization (OPSO) algorithm. The proposed algorithm relies on incorporating an optical-based updating mechanism. To evaluate the performance of the proposed model, two experiments were conducted. In the first experiment, the main objective is to compare the proposed OPSO and traditional PSO for handling global optimization functions. From this experiment, we demonstrated the superiority of OPSO, which achieved the smallest fitness value and converged faster than PSO. Moreover, the second experiment assessed the applicability of the developed OPSO model for detecting dyslexia. The results of OPSO have been compared with other well-known feature selection techniques. The results illustrated the high ability of the OPSO to reduce the number of features and increase the prediction performance. Finally, the results of the two experiments indicate that OPSO not only enhanced optimization accuracy and strong stability compared to the standard PSO algorithm but also possesses superior discriminative ability in dyslexia detection.

Based on the promising results, the proposed OPSO can be applied in future works in different applications. For example, apply it to enhance Autism detection and other disability diseases. Moreover, the use of the optical mechanism integrated into OPSO to solve other optimization problems outside the medical field could also be explored.