Research report: Deep learning based optimization of personalized psychological counseling strategies

The model, termed GBFI–EDLSTM, combines an Encoder–Decoder Long Short-Term Memory network with a bio-inspired optimization algorithm called Grizzly Bear Fat-Increase Optimization, and achieves an accuracy of 0.936 in identifying the correct counseling approach from student conversations. The motivation behind the research stems from a fundamental limitation of existing counseling support systems: most are built on static or cross-sectional data and cannot adapt to the fluctuating emotional states that unfold across successive dialogue turns. Traditional machine learning methods such as support vector machines, random forests, and logistic regression treat each interaction as an isolated event, ignoring the sequential nature of a counseling conversation.

Deep learning approaches improve contextual learning but often focus on single turns without maintaining emotional continuity. Reinforcement learning-based systems, meanwhile, depend on predefined reward designs that may not correspond to genuine psychological outcomes. Zhang's framework was designed to overcome these shortcomings by integrating three components in a single pipeline: BERT-based contextual embeddings for semantic and emotional understanding, an Encoder–Decoder LSTM for modeling temporal dependencies across the conversation, and GBFI optimization for adaptive parameter tuning.

The study used the publicly available Student Psychological Counseling Dataset from Kaggle, which contains 4000 counseling encounter records for college students across 15 columns. Each row represents one counseling interaction, capturing the student's reported concern, emotional state, contextual background, counselor response, and the counseling strategy ultimately applied. The dataset combines structured categorical variables such as age, gender, year of study, GPA range, session type, and session duration with unstructured text fields containing the conversation and the counselor's reply.

The target variable is the counseling strategy label itself, making the task a multi-class classification problem rather than an open-ended response generation system. The data was split into training and testing subsets using both 70:30 and 80:20 ratios to ensure robust evaluation. Before feature extraction, the text data underwent a series of preprocessing steps designed to reduce lexical variation and remove noise.

Tokenization broke the unstructured counseling conversations into machine-readable units while preserving semantic and syntactic boundaries. Lemmatization normalized morphologically different words such as feeling, felt, and feels into their root forms, allowing the model to treat them as semantically equivalent. Stop-word removal filtered out common words like the, a, is, and are that contribute little to emotional or contextual meaning, and punctuation removal stripped away non-informative symbols.

These steps produced cleaner inputs for BERT-based embedding extraction, which used BERT-base with 768-dimensional embeddings to capture rich contextual representations of the counseling dialogues. The Encoder–Decoder LSTM architecture forms the core of the sequential modeling component. The encoder, an LSTM layer with 256 hidden units, processes the tokenized input sequence and produces hidden and cell state vectors that encode both the semantics and the emotional content of the conversation.

These final states are passed to the decoder, a second LSTM layer with the same number of hidden units, which generates the output based on the encoded context. The gates within each LSTM cell—the forget gate, input gate, and output gate—selectively retain useful previous emotional information, integrate new language and emotion-related data from the current input, and control the flow of information to the next state. This gating mechanism allows the model to maintain temporal emotional sequences across dialogue turns, which is essential for tracking how a student's psychological state evolves during a session.

GBFI, the optimization component, is a bio-inspired metaheuristic algorithm that fine-tunes the trainable parameters of the EDLSTM network. It operates in three phases. In the global exploration phase, candidate solutions search across the parameter space guided by multiple reference solutions rather than a single optimum, preventing premature convergence.

In the exploitation phase, the algorithm shifts toward the best areas discovered, refining the solutions through an adaptive step size that decreases across iterations. In the local search phase, a fine-tuning mechanism with an oscillatory adjustment factor polishes the most promising solutions to ensure accurate and emotionally appropriate outputs. The optimization objective is a composite loss function that combines sequence prediction error with an emotional alignment term, weighted by a coefficient λ, ensuring that the model optimizes both semantic correctness and emotional appropriateness simultaneously.

The model was implemented in Python using TensorFlow and Keras for the deep learning components and NLTK and spaCy for text preprocessing. Training was conducted for 50 epochs with a learning rate of 0.001, a batch size of 32, and the Adam optimizer, with dropout regularization at a rate of 0.3 applied to the recurrent layers. Categorical cross-entropy served as the primary classification loss function.

During training, Adam performed local gradient-based updates while GBFI searched globally for better parameter sets, preventing the network from settling into local optima. Stable convergence was achieved within 100 epochs, with training accuracy reaching 0.96 and validation accuracy stabilizing at 0.85, while training loss dropped from 1.9 to 0.12 and validation loss reached 0.25, indicating that overfitting was effectively controlled. The experimental results demonstrate strong performance across multiple validation procedures.

Using an 80:20 train-test split, the GBFI–EDLSTM model achieved an accuracy of 0.936, precision of 0.93, recall of 0.92, and an F1-score of 0.92. With a 70:30 split, the model recorded an accuracy of 0.928, precision of 0.92, recall of 0.90, and an F1-score of 0.91. Beyond classification metrics, the model was evaluated on therapeutic indicators: the PHQ-9 depression score decreased by 4.3 points under the 80:20 split and 4.1 points under the 70:30 split, and the dropout rate fell to 8.8 percent and 9.5 percent respectively.

User satisfaction reached 4.6 and engagement 4.8 on their respective scales. A fivefold stratified cross-validation confirmed these findings, yielding a mean accuracy of 0.935 with a standard deviation of just 0.0023, indicating consistent and reliable performance across different data partitions. Comparative analysis against baseline methods revealed the superiority of the proposed framework.

The MPHI-Trans model, a transformer-based multimodal temporal framework, achieved an accuracy of 0.854 on the same data, while the RL-DSF reinforcement learning-based decision support framework reached 0.895. The proposed model outperformed both in precision, recall, F1-score, PHQ-9 reduction, dropout rate, satisfaction, and engagement. The RL-DSF approach reduced PHQ-9 scores by less than 3 points and had a dropout rate of 12.4 percent, while MPHI-Trans showed a dropout rate of 16.0 percent.

An ablation study confirmed that each component contributes incrementally: preprocessing and BERT embeddings improve performance over a scratch-LSTM baseline, the Encoder–Decoder architecture adds further gains, and GBFI optimization delivers the highest scores across all metrics. Despite the strong results, the study acknowledges several limitations. The model's effectiveness depends heavily on the quality and diversity of the training data, and its generalization to real-world clinical populations remains untested.

Only textual conversational data was used, excluding multimodal signals such as speech prosody or facial expression that could enrich emotional understanding. The GBFI optimization adds computational complexity to the training process, and the model has not yet undergone clinical validation in actual counseling settings. The framework is evaluated on an offline dataset and does not currently support online learning, longitudinal adaptation, or real-time deployment.

Future directions include multimodal input integration, real-time execution within chatbot platforms, cross-cultural validation, explainable AI for transparency in counseling decisions, and privacy-preserving data management to address the fairness and confidentiality concerns that remain significant obstacles in AI-enabled mental health systems.

Subject of Research: Deep learning based optimization of personalized psychological counseling strategies

Article Title: Deep learning based optimization of personalized psychological counseling strategies

Article References: Zhang, M. (2026). Deep learning based optimization of personalized psychological counseling strategies. Discover Artificial Intelligence, 6(1), Article 1271. https://doi.org/10.1007/s44163-026-01877-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01877-x

Keywords: Deep, learning, based, optimization, personalized, psychological, counseling, strategies, scientific research