Adaptive learning rate optimized hybrid model for early identification of children with autism

Architecture

The architecture of the ALROH model for ASD prediction is presented in Fig. 2. The process begins with preprocessed datasets, which are initially stored in .arff format for compatibility with standard preprocessing tools (e.g., WEKA). To simulate a FL environment, this dataset is partitioned into multiple client subsets, each representing an independent data holder. Instead of centralizing the raw data, each client subset is trained locally, and the learned parameters are aggregated on a central server using the FL mechanism. This setup ensures that model training simulates privacy-preserving and distributed learning rather than direct centralized training. Each client employs a CNN–LSTM model for feature extraction and sequential behavior analysis. The CNN identifies key patterns and dependencies in the behavioral and sensory data through convolution operations, ReLU activation, pooling, and fully connected layers. The extracted feature maps are then passed to the LSTM, which is well-suited for modeling the sequential dependencies that characterize ASD-related behavioral patterns. To enhance the overall predictive accuracy, a DFO algorithm is applied to optimize the hyperparameters and weights of the CNN–LSTM model. The DFO utilizes swarm intelligence to dynamically adjust network parameters, minimize prediction errors, and enhance convergence across the federated nodes. Finally, the aggregated model parameters are forwarded to the output layer, which generates a binary classification result: ASD-positive (1) or ASD-negative (0). This approach facilitates early detection of ASD, enabling timely interventions and personalized support for children at risk. Thus, the ALROH architecture integrates simulated FL with deep learning models (CNN and LSTM) and optimization techniques (DFO) to deliver an accurate, scalable, and privacy-aware framework for ASD prediction.

CNN for feature extraction

To handle ASD tabular datasets in .arff or .csv format, the proposed method first preprocesses each record into a structured matrix representation, where rows correspond to ASD-related attributes (behavioral, demographic, or screening-question features) and columns correspond to individual samples. This transformation enables convolutional operations in CNNs to exploit local correlations between neighboring features. The input feature matrix $I$ is denoted in Eq. (1).

(1) $$I=\left\{x\left(m,n\right)|1\le m\le W,1\le n\le H\right\}.$$

In Eq. (1), $x\left(m,n\right)$ denotes ASD-related features at every row $m$ and column $n$, $W$ denotes the total number of features and $H$ denotes different samples of ASD. Although CNNs are conventionally applied to image data, recent studies (e.g., ConvTabNet (Borisov et al., 2022), TabCNN (Kulkarni, 2024)) demonstrate that 1D/2D convolution can effectively capture spatial feature interactions in tabular datasets (Alenizy & Berri, 2025). Thus, a modified 1D-CNN is adopted here, where convolution kernels slide across the feature dimension to extract higher-order interactions among attributes. Therefore, every feature vector is transformed using a convolution filter to capture the local dependencies present in each sample; that is, every neuron in the convolution layer analyzes every feature of the input ASD data. In the convolution layer, the process of convolution is to extract spatial patterns through a filter/kernel $K$ having size $w_k\times h_k$. Using this, the feature-map $Y_i$ is constructed using the following operation presented in Eq. (2).

(2) $$Y_i={\left(I\otimes K\right)}_{m,n}=\sum _{u=-w_k}^{w_k}\sum _{v=-h_k}^{h_k}{K}_{u,v}\cdot I_{m+u,n+v}.$$

In Eq. (2), $K_{u,v}$ denotes kernel weights at position $\left(u,v\right)$ and summation denotes the overall kernel-window. Further, the convolution slides over $I$ having stride $S_k$, which defines the step-size movement, and zero-padding is applied to ensure dimensional consistency. Further, for every convolution-layer $l$, the feature map $Y_i^{\left(l\right)}$ is evaluated using Eq. (3).

(3) $$Y_i^{\left(l\right)}=f\left(B_i^{\left(l\right)}+\sum _{j=1}^{f^{\left(l-1\right)}}{K}_{i,j}^{\left(l\right)}\ast Y_j^{\left(l-1\right)}\right).$$

In Eq. (3), $B_i^{\left(l\right)}$ denotes a bias matrix for every $i^{th}$ feature map, $K_{i,j}^{\left(l\right)}$ denotes a convolutional kernel between feature maps $j$ of the previous layer and feature-maps $i$ of the current layer and $f\left(x\right)$ denotes Rectified Linear Unit (ReLU) activation function, as presented in Eq. (4).

(4) $$f\left(x\right)=\mathrm{m}\mathrm{a}\mathrm{x}\left(0,x\right).$$

Hence, using Eq. (3), the output feature map using a convolution layer for position $\left(m,n\right)$ is evaluated using Eq. (5).

(5) $$Y_{i,m,n}^{\left(l\right)}=f\left(B_{i,m,n}^{\left(l\right)}+\sum _{j=1}^{f^{\left(l-1\right)}}\sum _{u=-w_k}^{w_k}\sum _{v=-h_k}^{h_k}{K}_{i,j,u,v}^{\left(l\right)}\ast Y_{j,m+u,n+v}^{\left(l-1\right)}\right).$$

Using Eq. (5), the convolution layer enables hierarchical feature learning. After the convolution layer, the extracted features are passed on to the pooling layer, where they are reduced in complexity while retaining the dominant features. In the pooling layer, by utilizing max-pooling with having window size $w_p\times h_p$ and stride $S_p$, the pooling layer is evaluated as presented in Eq. (6).

(6) $$P{\left(Y_i^{\left(l\right)}\right)}_{m,n}=max{\left(Y_i^{\left(l\right)}\right)}_{m\le u\le m+w_p,n\le v\le n+h_p}.$$

In Eq. (6), $P$ denotes the pooling process. Using Eq. (6), only the highest value is retained for every window. Hence, the output dimensions for the pooling layer are defined using Eq. (7).

(7) $$d_p=\lfloor {\displaystyle \frac{W-w_k}{S_p}+1}\rfloor \times \lfloor {\displaystyle \frac{H-h_k}{S_p}+1}\rfloor .$$

After multiple convolution layers and a pooling layer, the output is flattened into a vector form denoted as $z$, which is then passed on to the Fully Connected Layer (FCL), where the following operation is performed as presented in Eq. (8).

(8) $${\left(Y_i^{\left(l\right)}\right)}_{m,n}=f\left(\sum _{p=1}^{f_1^{\left(l-1\right)}}\sum _{q=1}^{f_2^{\left(l-1\right)}}\sum _{r=1}^{f_3^{\left(l-1\right)}}{w}_{i,p,q,r}^{\left(l\right)}{\left(Y_j^{\left(l-1\right)}\right)}_{q,r}\right).$$

In Eq. (8), $f_1^{\left(l-1\right)}$, $f_2^{\left(l-1\right)}$, $f_3^{\left(l-1\right)}$ denotes total feature maps extracted using FCL, $w_{i,p,q,r}^{\left(l\right)}$ denotes weights that connect every row $m$ and column $n$, i.e., $\left(m,n\right)$ for every feature-map for $l^{th}$ layer for every position $\left(q,r\right)$ in feature-map $p$ present in $\left(l-1\right)$ Layer. To avoid overfitting in FCL, a dropout regularization function was applied, which reduced the overall number of neurons. Moreover, softmax was applied to convert the raw output features extracted using a modified CNN process into a binary class probability distribution using Eq. (9).

(9) $$p\left(y=i|I\right)={\displaystyle \frac{\exp \left(W_i^{T}I\right)}{{\sum }_{j=1}^C\exp \left(W_j^{T}I\right)}.}$$

In Eq. (9), $p\left(y=i|I\right)$ denotes the probability of the given input feature matrix $I$ belongs to a class $i$, i.e., 1 or 0 for ASD-positive or ASD-negative, respectively, $I$ denotes the input feature matrix, which is extracted from the CNN and processed by LSTM, as discussed in the next section, $W_i^T$ denotes weight vector for class $i$ which maps to extracted feature to its respective class $i$, $W_j^{T}I$ denotes the dot product of the weight vector $W_i$ and $I$, producing a scalar score for the class $i$, $\exp \left(W_j^{T}I\right)$ denotes the applied exponential function for class-score and ${\sum }_{j=1}^C\exp \left(W_j^{T}I\right)$ denotes the normalization term, which ensures the probabilities sum for the class $C$. Once the CNN layer extracted the hierarchical feature embeddings from structured ASD records. Pooling is then applied to reduce dimensionality, followed by flattening into a feature vector, which is passed to the LSTM. The LSTM model, which handles sequential dependencies in the dataset, is discussed in detail in the next section.

LSTM model for ASD prediction

After extracting features using a modified CNN, the feature matrix is passed on to LSTM, which captures sequential patterns and temporal dependencies from ASD datasets for ASD prediction. The flattened CNN feature embeddings are sequentialized to exploit dependencies among attributes. While LSTMs are designed for temporal sequences, each ASD record is represented here as a pseudo-sequence of features, where the ordering of attributes enables the LSTM gates to capture inter-feature dependencies that a CNN alone cannot model. Additionally, the LSTM in this work is utilized to address the vanishing gradient problem and learn from long-term dependencies using gates that control data flow. For every extracted feature using the modified CNN from the input feature matrix $I$, the LSTM cell processes the input sequence using Eq. (10).

(10) $$I=\left(i_1,i_2,\dots ,i_t\right).$$

In Eq. (10), $i_t$ is the input for the time step $t$. The LSTM cell utilizes two states at every step. $t$, i.e., hidden-state $h_t$ and cell-state $c_t$. The $h_t$ contains data related to upcoming steps and $c_t$ stores long-term dependency. Further, the LSTM cell consists of three primary gates, forget-gate $f_t$, input-gate $i_t$ and output gate $o_t$. The $f_t$ decides which previous data has to be forgotten, $i_t$ determines which new data has to be stored and $o_t$ controls which data is passed on to the next time step. The $f_t$ decides which data has to be forgotten or retained using previous cell-state $c_{t-1}$. Moreover, the $f_t$ uses input $x_t$ and previous $h_{t-1}$, applies weight-transformation and passes to the sigmoid-activation function as presented in Eq. (11).

(11) $$f_t=\sigma \left(W_f{x}_t+U_f{h}_{t-1}+b_f\right).$$

In Eq. (11), $f_t$ has activation values $0$ (close-forget past data) and $1$ (retain past data), $W_f$ denotes weight-matrix for $x_t$, $U_f$ denotes weight-matrix for $h_{t-1}$, $b_f$ is a bias term. The $\sigma \left(x\right)$ is a sigmoidal activation function, which is evaluated using Eq. (12).

(12) $$\sigma \left(x\right)={\displaystyle \frac{1}{1+e^{-x}}.}$$

The $i_t$ consists of two processes, i.e., input-gate activation $i_t$ and candidate cell-state ${\tilde{c}}_t$ which are evaluated using Eqs. (13) and (14), respectively.

(13) $$i_t=\sigma \left(W_i{x}_t+U_i{h}_{t-1}+b_i\right)$$

(14) $${\tilde{c}}_t=\tanh \left(W_c{x}_t+U_c{h}_{t-1}+b_c\right).$$

In Eqs. (13) and (14), $i_t$ controls how much new data enters and $W_i$, $W_c$, $U_i$, $U_c$, $b_i$ and $b_c$ are weights and bias terms, ${\tilde{c}}_t$ denotes candidate-memory content, which helps in identifying which new data has to be added and $\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(\mathrm{x}\right)$ is a hyperbolic-tangent activation function, which is evaluated using Eq. (15).

(15) $$\tanh \left(x\right)={\displaystyle \frac{e^x-e^{-x}}{e^x+e^{-x}}.}$$

Moreover, the $c_t$ is updated utilizing $f_t$ and $i_t$ using Eq. (16).

(16) $$c_t=f_t\odot c_{t-1}+i_t\odot {\tilde{c}}_t.$$

In Eq. (16), $c_t$ denotes updated cell state, $\odot$ denotes element-wise multiplication, $f_t\odot c_{t-1}$ ensures past memory is kept and $i_t\odot {\tilde{c}}_t$ ensures new data is added. Finally, the $o_t$ determines which $c_t$ has to be sent as $h_t$. Similar to $i_t$, $o_t$ also consists of two processes, first where output-gate activation is evaluated and then $h_t$ is updated. The output-gate activation is evaluated using Eq. (17).

(17) $$o_t=\sigma \left(W_o{x}_t+U_o{h}_{t-1}+b_o\right).$$

In Eq. (17), $o_t$ determines output strength and $W_o$, $U_o$ and $b_o$ are weights and bias terms. Further, the $h_t$ is updated using Eq. (18).

(18) $$h_t=o_t\odot tanh\left(c_t\right).$$

In Eq. (18), $h_t$ has relevant data from $c_t$ and $tanh$scales $c_t$ between −1 and 1 for a stable learning process. After the learning process using LSTM, the ASD prediction is performed. This work presents the DFO approach for optimizing hyperparameters of the CNN-LSTM model, which is discussed in detail in the next section.

Dragon-fly optimization

After extracting features from a CNN and utilizing sequential learning with LSTM, this work presents a DFO approach that optimizes the hyperparameters of CNN-LSTM to enhance accuracy and minimize loss. The CNN–LSTM framework involves critical hyperparameters (learning rate, number of CNN filters, LSTM hidden units, dropout rate, batch size). To avoid manual tuning, the DFO metaheuristic is employed. Each dragonfly represents a candidate hyperparameter set, and its position is updated by balancing exploration (global search) and exploitation (local refinement). DFO is a swarm-intelligent approach inspired by the swarm behaviour of dragonflies. The dragonflies move in groups during the exploration phase and remain in a static state during the exploitation phase, based on five key forces. These forces help optimize hyperparameters such as batch size, learning rate, LSTM total units, dropout rate, and others. In the DFO approach, every dragonfly denotes a candidate approach $S_i$ in hyperparameter searching space. The position update of dragonfly is evaluated using Eq. (19).

(19) $$S_i^{t+1}=S_i^t+w_m{M}_i+w_s{S}_i+w_a{A}_i+w_c{C}_i+w_e{E}_i.$$

In Eq. (19), $S_i^t$ denotes dragonfly position $i$ at time $t$ (current hyper-parameter values), $M_i$ denotes memory-based movement (previous best position), $S_i$ denotes separation, i.e., avoid collision or overcrowding, $A_i$ denotes alignment, which helps to synchronize velocity with respect to its other dragonfly neighbors, $C_i$ denotes cohesion, i.e., the dragonfly moves towards center of its neighboring solutions, $E_i$ denotes attraction towards the best approach found during optimization, $w_m$, $w_s$, $w_a$, $w_c$ and $w_e$ denote weight coefficients that control the impact of every movement factor. The $M_i$ is evaluated as Eq. (20).

(20) $$M_i=S_i^{t-1}.$$

The Eq. (20) ensures every dragonfly remembers its best solution and returns to them if a new position does not improve fitness. Further, the $S_i$ is evaluated using Eq. (21).

(21) $$S_i=-\sum _{j=1}^N\left(S_i-S_j\right).$$

The Eq. (21) avoids collision with other dragonflies by repelling dragonflies that are too close and $N$ denotes the total number of neighboring solutions. The $A_i$, $C_i$ and $E_i$ are evaluated using Eqs. (22), (23) and (24), respectively.

(22) $$A_i={\displaystyle \frac{1}{N}\sum _{j=1}^N{V}_j}$$

(23) $$C_i={\displaystyle \frac{1}{N}\sum _{j=1}^N{S}_j-S_i.}$$

(24) $$E_i=S_g-S_i.$$

In Eq. (22), the $A_i$ ensures the dragonflies’ velocity is the same as its neighbors, in Eq. (23) ensures the movement of dragonflies toward the swarm center and in Eq. (24), every dragonfly moves towards the best global solution $S_g$ found so far. This position update approach provides a framework for efficiently exploring and exploiting the search space, thereby enhancing hyperparameters for ASD prediction. Furthermore, the fitness function is used to evaluate every dragonfly’s solution. Hence, the accuracy and loss of the CNN-LSTM model for ASD prediction are evaluated using a fitness function guiding optimization in Eq. (25).

(25) $$F=\alpha \cdot Accuracy-\beta \cdot Loss.$$

In Eq. (25), $F$ denotes fitness-score, $Accuracy$ denotes the prediction accuracy of the CNN-LSTM model, $Loss$ denotes loss attained during cross-entropy and $\alpha$ and $\beta$ are balancing coefficients. Through iterative position–velocity updates, DFO avoids premature convergence and identifies hyperparameters that maximize ASD prediction accuracy. In DFO, every candidate solution (dragonfly) is tested utilizing Eq. (25), and the best-performing hyperparameters are retained. Further, every dragonfly updates its respective velocity to control convergence speed using Eq. (26).

(26) $$V_i^{t+1}=w\cdot V_i^t+r_1\left(P_i-S_i\right)+r_2\left(G-S_i\right).$$

In Eq. (26), $V_i^t$ is the velocity of a dragonfly $i$ at time $t$, $w$ denotes inertia weight, $P_i$ denotes best local solution found is dragonfly, $G$ denotes the best global solution found by Dragonfly and $r_1$, $r_2$ denotes random coefficients, which ensure exploration and avoid premature convergence. Moreover, in Eq. (26), $w\cdot V_i^t$ keeps momentum using the previous velocity, $r_1\left(P_i-S_i\right)$ moves toward a best local solution and $r_2\left(G-S_i\right)$ moves towards the best global solution. This velocity update process allows dragonflies to adapt dynamically, reduce search stagnation, and find the best solution. The complete hyper-tuning process for the CNN-LSTM model is presented in Algorithm 1 below.

Using the DFO in the CNN-LSTM model, the approach avoids local minima, balances random exploration, handles class imbalance, provides focused learning, and finds the best optimal values for hyperparameter tuning. The overall ALROH model achieves higher prediction accuracy using hyperparameter tuning, which is discussed in detail in the next section.

System requirements and performance metrics

To evaluate the performance of the proposed ALROH model, experiments were conducted on a system equipped with 16 GB of RAM and an Intel Core i7 processor. The model was implemented using Python and executed within the Anaconda environment, which was installed on a Solid-State Drive (SSD) to enhance computational speed and efficiency. Using an SSD helped reduce data retrieval time, thereby optimizing the training and evaluation process. Standard performance metrics were also employed to assess the effectiveness and reliability of the ALROH model. These included accuracy, precision, recall, and F-score, widely used in ML to evaluate predictive models. The mathematical formulations of these metrics are presented in Eqs. (27), (28), (29), and (30), ensuring a comprehensive assessment of the model’s predictive capabilities.

(27) $$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(28) $$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(29) $$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(30) $$F\text{-}score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}.}$$

In Eqs. (27), (28), (29) and (30), $TP$ and $FP$ denotes true-positive and false-positive, and $TN$ and $FN$ denotes true-negative and false-negative.

Datasets

In this work, three publicly available ASD datasets were utilized to evaluate the performance of the proposed ALROH model. The goal was to assess the model’s predictive capabilities across different datasets and ensure its generalizability in ASD detection. The first dataset was the children dataset, collected by Thabtah (2018). This dataset is openly accessible and can be downloaded from UCI Machine Learning Repository (2025). It contains 292 samples, with 141 ASD-positive cases and 151 non-ASD cases. Additionally, Thabtah (2018) also collected a toddler dataset, which is publicly available and can be accessed from Kaggle.com (2025). This dataset comprises 1,053 samples, where 728 cases were identified as ASD-positive, while 325 cases were classified as non-ASD, referred to as the toddler 1 dataset in this work. Furthermore, another toddler dataset was considered, which was collected by Alkahtani et al. (2023). This dataset is openly accessible and can be downloaded from Kaggle.com (2022). It comprises a total of 506 samples, consisting of 341 ASD cases and 165 non-ASD cases, which is referred to as the toddler 2 dataset in this work. With respect to population context, the datasets employed in this study were originally collected by Thabtah (2018) and Alkahtani et al. (2023) using standardized screening questionnaires (such as Autism Spectrum Quotient—AQ-10) and behavioral indicators. The participants were toddlers and children flagged for screening in educational and healthcare contexts. These datasets do not explicitly categorize children by ASD severity levels or therapy participation; instead, they label children as “ASD-positive” or “non-ASD” based on screening thresholds. Therefore, the data reflect screening-level detection rather than detailed clinical severity assessments or therapy outcomes. This makes the datasets particularly suitable for evaluating early-detection models such as ALROH, which aim to assist schools in identifying at-risk children for further professional evaluation. Thus, enhance the robustness of the evaluation, a combined toddler dataset was created by merging the datasets collected by Thabtah (2018) and Alkahtani et al. (2023). The merged dataset comprised 1,559 samples, including 1,069 ASD-positive cases and 490 non-ASD cases, which is referred to as the merged toddler dataset in this work. Finally, a comprehensive dataset was generated by merging the children and toddler datasets collected by Thabtah (2018) and the toddler dataset collected by Alkahtani et al. (2023). The final merged dataset consisted of 1,851 samples, comprising 1,210 ASD cases and 641 non-ASD cases, which is referred to as the merged dataset in this work. Five different datasets were used to rigorously evaluate the ALROH model. By considering multiple datasets with varying distributions of ASD and non-ASD cases, this study aimed to ensure the reliability, effectiveness, and generalization capability of the proposed model across different age groups and data distributions.

Research questions and hypotheses: The empirical evaluation of ALROH was guided by three research questions (RQ): RQ1: Can federated learning (FL) enhance the generalizability of ASD prediction models by securely integrating data from multiple sources? RQ2: Does a hybrid CNN–LSTM architecture capture both spatial and sequential behavioral patterns more effectively than conventional models? RQ3: To what extent does Dragonfly Optimization (DFO) improve hyperparameter tuning, convergence speed, and robustness under class-imbalanced conditions? Based on these, the following hypotheses were tested: (i) FL improves model generalization across datasets; (ii) CNN–LSTM outperforms baseline ML/DL approaches in accuracy and recall; (iii) DFO enhances predictive stability and convergence efficiency compared to non-optimized baselines.

Children dataset

The performance of the proposed ALROH model was thoroughly evaluated using the ASD children dataset, and the results are illustrated in Fig. 3. The model demonstrated exceptional predictive capability across all standard performance metrics. Specifically, the ALROH model achieved 99.23% accuracy, indicating its high reliability in correctly classifying ASD and non-ASD cases. Additionally, the model achieved a precision of 99.23%, indicating its ability to minimize false positives while accurately identifying ASD cases. A 99.21% recall further highlights the model’s effectiveness in correctly detecting ASD cases, ensuring that very few ASD-positive instances were misclassified. Finally, the 99.22% F-score reflects the model’s balanced performance, demonstrating its robustness in predicting ASD.

Toddler dataset 1

The performance evaluation of the ALROH model on the ASD Toddler 1 dataset is presented in Fig. 4, demonstrating the model’s effectiveness in accurately predicting ASD in toddlers. The ALROH model achieved 99.36% accuracy, indicating strong predictive performance and reliability in classifying ASD and non-ASD cases. Additionally, the model achieved 99.32% precision, ensuring minimal false positives while accurately identifying ASD-positive instances. A 99.33% recall further highlights the model’s ability to accurately detect ASD cases, thereby minimizing false negatives (FNs) and improving early diagnosis. Moreover, a 99.34% F-score reflects a balanced performance, signifying the overall robustness of the model.

Toddler dataset 2

The performance evaluation of the ALROH model on the ASD Toddler 2 dataset is presented in Fig. 5, showcasing its high efficiency in predicting ASD among toddlers. The ALROH model achieved 99.28% accuracy, indicating its strong capability in distinguishing ASD and non-ASD cases with minimal errors. The model also demonstrated 99.26% precision, ensuring that FPs were kept to a minimum while accurately identifying ASD cases. Additionally, the 99.27% recall highlights the model’s effectiveness in accurately detecting ASD cases, thereby minimizing FNs and enhancing early diagnosis. The 99.27% F-score further validates the model’s balanced performance, demonstrating its robustness in classifying ASD.

Merged toddler dataset

The performance evaluation of the ALROH model on the ASD merged toddler dataset is presented in Fig. 6, demonstrating its effectiveness in handling a larger and more diverse dataset. The ALROH model achieved 98.61% accuracy, demonstrating its robustness and generalization capabilities across various toddler ASD datasets. The 98.6% precision reflects the model’s ability to classify ASD cases while minimizing FPs correctly. Additionally, 98.6% recall highlights the model’s efficiency in identifying true ASD cases, ensuring minimal FNs. The 98.6% F-score further confirms the model’s balanced performance, making it a reliable approach for ASD prediction in toddlers. The slightly lower accuracy compared to individual toddler datasets suggests that the merged dataset introduces greater variability. Yet, ALROH still maintains high performance, showcasing its adaptability and effectiveness in predicting ASD across different datasets.

Merged dataset

The performance evaluation of the ALROH model on the ASD merged dataset is presented in Fig. 7, highlighting its strong predictive capabilities across a diverse dataset that includes both children and toddlers. The ALROH model achieved 99.34% accuracy, demonstrating its ability to generalize well across different age groups and dataset variations. The 99.32% precision indicates the model’s efficiency in correctly identifying ASD cases while minimizing FPs. Furthermore, 99.31% recall reflects the model’s effectiveness in detecting ASD cases, ensuring that very few actual ASD cases are misclassified. The 99.31% F-score further validates the model’s balanced performance. These high-performance metrics indicate that ALROH can effectively handle larger and more diverse datasets, making it a promising tool for early ASD detection in both children and toddlers, thereby improving screening accuracy and facilitating timely interventions. Further, in the next section, the ALROH model is compared with existing approaches discussed in the literature survey.

Ablation study

To evaluate the contribution of each component in the ALROH model, an ablation study was conducted using the merged toddler dataset. The study assessed the model’s performance under four conditions: (1) ALROH without DFO, using standard gradient descent for optimization; (2) ALROH without LSTM, replacing it with a fully connected layer for prediction; (3) ALROH without CNN, using raw input features directly fed to LSTM; and (4) ALROH without FL, trained on a centralized dataset. The results, summarized in Table 2, demonstrate that the full ALROH model (with CNN, LSTM, DFO, and FL) achieved the highest accuracy (98.61%), precision (98.6%), recall (98.6%), and F-score (98.6%). Removing DFO reduced accuracy to 96.5%, indicating its role in optimizing hyperparameters. Replacing the LSTM with a fully connected layer decreased accuracy to 95.8%, highlighting the importance of the LSTM in capturing sequential patterns. Omitting CNN led to a significant drop in accuracy to 94.2%, underscoring its effectiveness in feature extraction. Finally, training without FL reduced accuracy to 97.1%, suggesting that FL enhances generalization through decentralized data. These findings confirm that each component is integral to ALROH’s superior performance.

Comparative study

This section discusses the results of the ALROH model and compares them with existing approaches. The results are compared in Table 3, which highlights the superiority of the proposed ALROH model over existing approaches. The ALROH model achieved 99.23% accuracy, outperforming previous models, including SVM (98%) from Farooq et al. (2023), RF (95.9%) from Pae & Pae (2024), L-R (96.23%) from Reghunathan et al. (2024), and LR (94.3%) from Bawa et al. (2024). Additionally, ALROH demonstrated significantly higher precision (99.23%), recall (99.21%), and F-score (99.22%), ensuring better overall performance in accurately predicting ASD cases. In contrast, the SVM model from Farooq et al. (2023) had 44% recall, indicating a high number of FNs, which is critical in ASD prediction as undiagnosed cases may not receive timely intervention. Similarly, while RF from Pae & Pae (2024) and LR from Reghunathan et al. (2024) yielded competitive results, their accuracy and recall values were lower than those of ALROH, suggesting a slightly reduced ability to classify ASD cases correctly. Furthermore, the LR model from Bawa et al. (2024) had the lowest accuracy (94.3%), making it the least effective among the compared approaches.

Furthermore, the comparative study on the ASD toddler dataset is presented in Table 4, demonstrating the ALROH model’s effectiveness compared to other state-of-the-art approaches. The ALROH model achieved 99.36% accuracy, outperforming models such as QDA-PCA (94%) from Alshammari et al. (2024) and Stacked Model (99.14%) from Das et al. (2025), while showing competitive performance against CNN-LSTM-PSO (99.64%) from Priyadarshini (2023), RF-XGB (99%) from Hajjej et al. (2024), and XGB 2.0 (99%) from Aldrees et al. (2024). In addition to its high accuracy, ALROH also demonstrated better precision (99.32%), recall (99.33%), and F-score (99.34%), ensuring a balanced performance in ASD prediction. Notably, while CNN-LSTM-PSO from Priyadarshini (2023) achieved a slightly higher accuracy of 99.64%, its precision (96%), recall (94%), and F-score (91%) were significantly lower than those of ALROH, indicating that it had more false positives and false negatives. On the other hand, QDA-PCA from Alshammari et al. (2024) had the lowest accuracy (94%) and an F-score of only 87%, suggesting weaker overall predictive performance. Moreover, compared to RF-XGB from Hajjej et al. (2024) and XGB 2.0 from Aldrees et al. (2024), both of which achieved 99% accuracy, the ALROH model still performed better in terms of precision and recall, making it more reliable and effective for ASD prediction in toddlers. The Stacked Model from Das et al. (2025) combined multiple classifiers and achieved strong results (99.14% accuracy, 98.76% precision, 99.37% recall, and 99.07% F-score), but its computational complexity was significantly higher than that of ALROH.

Practical Implications of Findings. ALROH’s superior performance is attributed to its hybrid DL framework, which integrates CNN for feature extraction, LSTM for sequential learning, and DFO for hyperparameter tuning. This combination enables better generalization and robustness in predicting ASD, making ALROH a highly efficient and effective model for early ASD detection in toddlers. Beyond achieving superior predictive performance, the results of ALROH carry important implications for practice. For parents, the availability of accurate and privacy-preserving screening tools may help raise early awareness of ASD-related behaviors, reducing delays in seeking professional evaluation and therapy. For teachers, the system can serve as a supportive decision-making aid, enabling early identification of students who may require differentiated instructional approaches or referrals to specialists. For school administrators and policymakers, ALROH demonstrates the feasibility of deploying federated learning frameworks that allow institutions to collaborate without compromising student privacy. This could guide the creation of national screening initiatives and inclusive education policies, ensuring equitable access to early interventions.

While the current datasets primarily focus on screening outcomes (ASD-positive or non-ASD), they do not include detailed socio-economic or clinical contexts, such as family education level, economic background, age stratification, or therapy history. Such variables could provide deeper insights into risk factors and strengthen the objectivity of ASD prediction frameworks. As part of future research, incorporating these variables into federated learning pipelines would allow models such as ALROH to capture both behavioral patterns and contextual risk factors, thus enhancing their relevance for educational planning and social policy design. In the next section, the conclusion, along with a detailed discussion of future work, is presented.