Enhancing Psychologists' Understanding through Explainable Deep Learning Framework for ADHD Diagnosis

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that is challenging to diagnose and requires advanced approaches for reliable and transparent identification and classification. It is characterized by a pattern of inattention, hyperactivity and impulsivity that is more severe and more frequent than in individuals with a comparable level of development. In this paper, an explainable framework based on a fine-tuned hybrid Deep Neural Network (DNN) and Recurrent Neural Network (RNN) called HyExDNN-RNN model is proposed for ADHD detection, multi-class categorization, and decision interpretation. This framework not only detects ADHD, but also provides interpretable insights into the diagnostic process so that psychologists can better understand and trust the results of the diagnosis. We use the Pearson correlation coefficient for optimal feature selection and machine and deep learning models for experimental analysis and comparison. We use a standardized technique for feature reduction, model selection and interpretation to accurately determine the diagnosis rate and ensure the interpretability of the proposed framework. Our framework provided excellent results on binary classification, with HyExDNN-RNN achieving an F1 score of 99% and 94.2% on multi-class categorization. XAI approaches, in particular SHapley Additive exPlanations (SHAP) and Permutation Feature Importance (PFI), provided important insights into the importance of features and the decision logic of models. By combining AI with human expertise, we aim to bridge the gap between advanced computational techniques and practical psychological applications. These results demonstrate the potential of our framework to assist in ADHD diagnosis and interpretation.

💡 Research Summary

The paper presents an explainable artificial intelligence (XAI) framework for diagnosing Attention Deficit Hyperactivity Disorder (ADHD) that combines a fine‑tuned hybrid deep neural network–recurrent neural network (HyExDNN‑RNN) with feature selection and interpretability techniques. Using the publicly available ADHD200 dataset, which contains demographic, clinical, and imaging‑related variables for 472 participants, the authors first preprocess the data by imputing missing values with column means, encoding categorical fields, and normalizing all features. They then apply Pearson correlation analysis, discarding features with a correlation coefficient above 0.95 to reduce redundancy, thereby retaining the most informative variables such as ADHD Index, Inattentive Score, IQ measures, and medication status.

The core classification model, HyExDNN‑RNN, consists of two dense layers (256 and 128 units with ReLU activation), dropout regularization (0.5), two recurrent layers (64 and 32 units, the first returning sequences), another dropout layer, a dense layer of 16 units, and a final softmax output layer with four neurons for multi‑class prediction. The model is compiled with the Adam optimizer and categorical cross‑entropy loss. In addition to this deep model, the authors train several conventional machine‑learning classifiers (Random Forest, Decision Tree, Extreme Gradient Boosting) and other deep variants (plain DNN, LSTM, LSTM‑GRU, LSTM‑RNN) for comparative evaluation.

Experimental results show that the HyExDNN‑RNN achieves an F1 score of 99 % on the binary ADHD vs. control task and 94.2 % on the four‑class problem (typical developing children, ADHD‑combined, ADHD‑inattentive, ADHD‑hyperactive/impulsive). These figures substantially outperform previously reported CNN‑based approaches (≈68–80 % accuracy) and the baseline machine‑learning models (≈80 % or lower on multi‑class).

To address the “black‑box” nature of deep learning, the study incorporates two XAI methods. SHapley Additive exPlanations (SHAP) provides local explanations by quantifying each feature’s contribution to an individual prediction, allowing clinicians to see why a specific subject is classified as a particular ADHD subtype. Permutation Feature Importance (PFI) offers a global perspective by measuring the impact on model performance when each feature’s values are randomly shuffled. Both techniques consistently highlight the importance of ADHD Index, Inattentive Score, and IQ‑related measures, aligning the model’s reasoning with established clinical markers. Visualizations of SHAP values and PFI rankings are presented as decision‑support tools for psychologists.

The authors discuss several limitations. The dataset exhibits severe class imbalance, especially the hyper‑active/impulsive class with only five instances, which may bias the model despite the high overall scores. Simple mean imputation for missing data could introduce subtle biases; more sophisticated approaches such as Bayesian imputation or multiple imputation are suggested for future work. Moreover, the current pipeline relies solely on tabular clinical variables; integrating raw neuroimaging data (e.g., resting‑state fMRI) in a multimodal architecture could further improve diagnostic accuracy and provide richer explanatory insights.

In conclusion, the paper delivers a robust, high‑performing ADHD diagnostic system that simultaneously satisfies the clinical demand for interpretability. By coupling a carefully engineered hybrid DNN‑RNN with SHAP and PFI, the framework not only achieves state‑of‑the‑art classification metrics but also equips psychologists with transparent, actionable explanations. This “AI‑in‑the‑loop” approach bridges the gap between advanced computational methods and practical mental‑health practice, paving the way for trustworthy AI‑assisted diagnosis in neurodevelopmental disorders.