Explainable Machine Learning for Predicting Judicial Decisions: A Case Study Across Civil and Criminal Courts

Abstract

The researchers proposed a hybrid EAI model to forecast outcomes in criminal and civil justice systems. The goal is to increase the predictive accuracy of judicial decision-making support systems while preserving their interpretation, fairness, and statistical validity by leveraging structured legal features and unstructured text representations. Case data, procedural history, characteristics of evidence, and legal textual narratives are included as case labels, making the problem a supervised learning problem. Several machine learning and deep learning models, including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, XG Boost, and the transformer-based BERT architecture, are used in the suggested framework. A feature representation is a merging of word embeddings, contextual embeddings and TF-IDF vectors. For cross-validation setting, accuracy, precision, recall, F1-score, and ROC-AUC measures are used for the evaluation of civil, criminal and hybrid legal datasets. The experimental results show that BERT has fairly good knowledge of legal semantics is fairly good for BERT, whereas XGBoost is more accurate in prediction. Judges are more predictable in civil trials than in criminal trials because of the greater discretion and more complex evidence. Introducing explainability, such as SHAP and LIME, enables interpretation of forecasts at the local and global levels through statutes, facts, and procedural history. The results are statistically validated using t-tests, the Wilcoxon signed-rank test, McNemar's test, and bootstrap confidence intervals, thereby providing credibility and validity. The paper presents a predictive paradigm of judicial decision-making across multiple domains that integrates prediction, stability, explainability and fairness, paving the way for responsible AI in legal decision-making processes and in computational legal research.