POLYSCORE: A Comparison of Accuracy.
Abstract
Using data collected under a mock crime scenario paradigm, four versions of the John Hopkins University Applied Physics Laboratory (APL) algorithm-based scoring system were evaluated for consistency in scoring accuracy. The four versions were: (a) PASS 2.0, (b) POLYSCORE 2.3, (c) POLYSCORE 2.9, and (d) POLYSCORE 3.0. The algorithm's rates of agreement/disagreement with ground truth were examined, and the same evaluations were made for the psychophysiological detection of deception (PDD) examiners who collected the data. The PDD examiners in this evaluation had an overall accuracy rate of 72.27% when compared to ground truth. The overall rate of accuracy generated by the algorithm (edited dataset) was: (a) PASS 2.0, 63.03%; (b) POLYSCORE 2.3, 67.72%; (c) POLYSCORE 2.9, 72.27%; and (d) POLYSCORE 3.0, 68.91%. With the inconclusive decisions eliminated, the recomputed accuracy rate for the PDD examiners was 79.63%, while each version of the algorithm was comparable (PASS 2.0, 78.95%; POLYSCORE 2.3, 79.21%; POLYSCORE 2.9 83.50%); POLYSCORE 3.0, 82.83%, with both POLYSCORE 2.9 and POLYSCORE 3.0 exceeding the examiners' level of accuracy. In addition to overall accuracy and accuracy based on the test format used, the effects of subjective manipulation of the data were discussed, and information was provided on the occurrence of decision reversals and statistical outliers.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1996
- Accession Number
- ADA313620
Entities
People
- N. J. Blackwell
Organizations
- United States Department of Defense