A major aspect of any forensic science discipline is the interpretation of results from the analysis of the evidence. As we incorporate artificial intelligence into our everyday lives, we ask whether it can assist us with forensic data interpretation to provide quick and reliable information that leads to a conclusion. In this study, three methods are investigated to determine if they could function like humans, are easily understood, and provide quality decision making that could be implemented in casework.
The scope of these three methods ranges from a complex machine learning application to an innovative application for data interpretation to a machine/analyst combined process. Convolutional neural networks are deep learning artificial neural networks that are typically used with visual imagery data. Here we use the total ion spectra of a fire debris sample. Images of the total ion spectra are the data input for the neural network. Another method is a fire debris analysis application that imitates the processes utilized by fire debris analysts in their data interpretation. The application utilizes multiple algorithms to collectively follow an analyst’s progression in interpreting data with the advantage of being able to process substantial amounts of data quickly. The third method combines the decision from an analyst’s interpretation with a decision from a machine learning method. Analysts perform their interpretation following a linear sequential unmasking workflow and incorporate a subjective logic approach that provides a score. The analyst workflow is designed to mitigate bias and provides a score representing the strength of evidence. The score is like a score resulting from a machine learning application. This allows the scores of the analyst and the machine learning application to be compared or pooled together.
The convolutional neural network emulates human processing of images which fire debris analysts perform in pattern recognition of total ion chromatograms of the extracted fire debris. The area under the receiver operating characteristic (ROC) curve is high at 0.86 for classification as to whether an ignitable liquid residue is present or not. The limitation of this method is its complexity that does not allow the user to easily understand how it makes decisions based on the image data. The fire debris analysis application is more of a tool for the forensic analyst rather than a machine learning method that provides a result. The subjective logic used in the analyst’s workflow combined with a machine learning application offers the best of both human and machine data interpretation. The area under the ROC curve ranges from 0.90 to 0.97 for three analysts and a support vector machine application. This method does require fire debris samples of known ground truth as to the presence or absence of ignitable liquid residue, a new interpretation workflow for human analysis, and validation of machine learning applications. The most difficult of these will be the acknowledgement that there is uncertainty in all decision-making processes. In this methodology the uncertainty of both human and machine decisions is calculated and considered in the final decision.
(Author abstract provided.)