Data-driven methods allow prediction of utility of DNA rework

In forensic DNA casework, complex trace samples may benefit from rework procedures involving additional PCR replicates and lowered analytical thresholds. These procedures may strongly increase or decrease the likelihood ratio (LR), simplifying decisions to be taken downstream. However, the procedures are resource-intensive. Decisions to pursue rework currently rely on laboratory guidelines and expert judgment rather than empirical evidence. This study aims to develop a model that predicts whether rework will yield a more decisive LR, defined as crossing the laboratory's reporting thresholds. To this end, we collected 2538 casework samples for which rework was performed and extracted features from the trace DNA profile and the initial LR calculation. An XGBoost classifier trained on these data demonstrated strong performance, achieving an area under the receiver operating characteristic curve (AUC ROC) of 91% and an accuracy of 82% on independent validation data (n = 121). However, a simple density-based model utilizing only a single feature, the initial log LR, performed nearly as good on the test set and comparable on the holdout set, with an AUC of 92% and an accuracy of 81%. When compared with assessments from 19 DNA experts on 23 cases, both models achieved an accuracy of 95%, compared to 72% for expert estimates. Experts frequently classified cases as uncertain (34%-66% success probability), whereas the models provided more decisive predictions. These results demonstrate that data-driven methods can enhance the efficiency and uniformity of forensic DNA workflows by providing evidence-based guidance for rework decisions, complementing expert judgment in resource allocation. In this application, the simpler model is preferred over more complex machine learning approaches due to its comparable performance and higher robustness and explainability.