Utilizing Text Mining, Data Linkage and Deep Learning in Police and Health Records to Predict Future Offenses in Family and Domestic Violence

George Karystianis, Rina Carines Cabral, Soyeon Caren Han, Josiah Poon, Tony Butler

Introduction

Family and Domestic Violence (FDV) is a global crisis with staggering social and economic costs—estimated at $22 billion annually in Australia alone. While police forces collect vast amounts of data, much of it remains "dark." Standard databases use "fixed fields" (like age or location), but the most nuanced information—descriptions of threats, victim injuries, and behavioral red flags—is buried in unstructured text narratives. Furthermore, while the link between mental health and FDV is well-documented, police and health records are rarely integrated. This paper addresses the challenge of "siloed" data by linking police narratives with actual clinical diagnoses to see if deep learning can identify patterns that human review might miss, ultimately aiming to provide law enforcement with better early-intervention tools.

Methods

The researchers employed a multi-stage computational pipeline to transform raw police and health data into predictive insights:

  1. Text Mining: A rule-based methodology was used to extract 44 abuse types, 17 injury types, and 126 mental health mentions from 492,393 police narratives.
  2. Data Linkage: Using probabilistic record linkage, police records were connected to NSW Ministry of Health data (Admitted Patient and Emergency Department records) to include actual clinical diagnoses for both victims and Persons of Interest (POIs).
  3. Modeling: The team structured the problem as a time-series experiment, using a window of up to three prior FDV events to predict the nature of the next event. They compared several architectures, including Multilayer Perceptron (MLP), sequential models (LSTM, Bi-LSTM, Bi-GRU), and BERT (Bidirectional Encoder Representations from Transformers) to capture contextual nuances.
  4. Interpretability: They utilized LIME (Local Interpretable Model-agnostic Explanations) to visualize which specific words in a narrative (e.g., "weapon," "punching") were driving the model’s predictions.

Datasets and Feature Extraction

The study utilized 492,393 FDV event narratives from the NSW Police Force (NSWPF) COPS Database (2005–2016) linked with NSW Ministry of Health records. The final preprocessed set consisted of 189,509 "windows" of sequential events.

Distribution of FDV offense types recorded in the police dataset.

As illustrated in Figure 3, "Hands-On" (physical) offenses were the most prevalent in the recorded events, followed by "Hands-Off" and "ADVO Breach" categories. The overlap in these categories necessitated a multilabel modeling approach.

Key Insights: The Power of Linkage

The feature ablation study revealed that models using the "ALL" dataset (fixed fields + text-mined features + linked health records) consistently outperformed those using only structured police data.

Frequency of features extracted across different offense categories via text mining and data linkage.

Figure 6 highlights the frequency of various feature groups extracted. The analysis yielded several critical observations:

  • Context vs. Structure: BERT was the champion for predicting ADVO breaches (breaching a court-issued protection order). These events are often defined by complex interpersonal dynamics that BERT’s contextual embeddings are uniquely suited to interpret.
  • Physical Violence Signal: Interestingly, MLP with categorical embeddings outperformed BERT for predicting "hands-on" (physical) offenses. This suggests that physical violence is more strongly correlated with discrete, high-signal features (like previous weapon use) than with subtle linguistic context.
  • The "Hands-off" Difficulty: Non-physical abuse (stalking, intimidation) remains the hardest to predict, likely because its features overlap significantly with the early stages of ADVO breaches.

Key Results

  • ADVO Breach Prediction: The BERT model achieved the best performance with 69.00% accuracy and a 66.76% ROC AUC.
  • Hands-on Offenses: The MLP model reached 65.95% accuracy and a high F1-score of 78.03%, indicating strong performance in identifying physical violence.
  • Hands-off Offenses: These were the most challenging, peaking at 60.72% accuracy using BERT.
  • Precision Focus: The researchers intentionally tuned models for high precision (up to 70% for physical offenses) to minimize the risk of "false positives" or unfair profiling.

Discussion and Future Work

While the results are promising, several hurdles remain. Police data only captures events where authorities were called, and current text mining tools struggle with entity resolution when multiple individuals are mentioned in one narrative. Furthermore, the authors stress that these "black box" models should assist, not replace, human expert judgment in justice settings. Future work aims to incorporate even more diverse data sources, such as GP visits and housing support records, to build a more holistic "360-degree" view of risk factors.

Summary updated

BibTeX
@article{karystianis2021utilizing,
  author = {George Karystianis and Rina Carines Cabral and Soyeon Caren Han and Josiah Poon and Tony Butler},
  title = {{Utilizing Text Mining, Data Linkage and Deep Learning in Police and Health Records to Predict Future Offenses in Family and Domestic Violence}},
  journal = {Frontiers Digit. Health},
  year = {2021},
  doi = {10.3389/fdgth.2021.602683},
}
Figure