An Intelligent Decision Support Framework for Malaysian Homicide Investigations Using the DeepConv-LSTM Model

Homicide investigations are among the most complex and resource-intensive tasks in criminal justice systems worldwide. In Malaysia, solving homicide cases often involves dealing with incomplete data, subjective witness statements, diverse crime scenes, and dynamic case developments. This research introduces an intelligent decision support framework for Malaysian homicide investigations using the DeepConv-LSTM model. The proposed hybrid architecture leverages convolutional layers to extract local patterns from structured and unstructured data, and LSTM layers to capture long-term temporal dependencies within investigative timelines. Datasets comprising official Malaysian police records and regional homicide case files from states such as Selangor, Johor, and Sabah were used. The model was trained using features like suspect demographics, crime location, weapon type, forensic reports, and sentiment from witness statements. The DeepConv-LSTM achieved a peak accuracy of 93.87%, outperforming traditional models such as SVM (82.31%) and Random Forest (86.74%). Additionally, the model showed strong F1-Score values, with an average of 91.63% across all regional datasets. This robust accuracy demonstrates the framework's potential to assist law enforcement in early detection of case solvability, prioritization of leads, and resource optimization. By combining deep learning with real-world legal and forensic data, this study contributes a novel methodology for digital forensics and intelligent criminal investigation.