Deception Detection: Using Machine Learning to Analyze 911 Calls.

Area of Science:

• Forensic Psychology
• Computational Criminology
• Artificial Intelligence in Law Enforcement

Background:

• Distinguishing false allegations from genuine reports in emergency calls is critical for resource allocation and justice.
• Previous methods for deception detection in 911 calls have limitations in accuracy and scalability.
• Machine learning offers a novel approach to analyze complex behavioral patterns in caller interactions.

Purpose of the Study:

• To evaluate the efficacy of machine learning, specifically the random forest model, in detecting deception in 911 calls.
• To identify specific behavioral cues that differentiate false allegation callers (FACs) from true report callers (TRCs).
• To assess the model's performance across different types of critical incident reports (homicide vs. missing person).

Main Methods:

• A dataset of 210 911 calls was compiled, comprising equal numbers of FACs and TRCs.
• Independent coders, blinded to deception status, analyzed calls for 86 distinct behavioral cues.
• A random forest model with k-fold cross-validation and repeated sampling was employed for classification.

Main Results:

• The random forest model achieved an overall accuracy of 68.2% in detecting deception across all 911 calls.
• Performance varied by report type: homicide reports yielded 71.2% accuracy, while missing person reports had 61.4% accuracy.
• Key discriminative cues included "Blames others," "Is self-dramatizing," and "Is uncertain and insecure."

Conclusions:

• Machine learning provides a viable tool for enhancing deception detection in emergency 911 calls.
• The identified behavioral cues offer valuable insights for training call-takers and investigators.
• Further research can refine models for improved accuracy in critical incident deception detection.