Development of a Deep-Learning-Based Computerized Scoring Algorithm for Polygraph Data

Development of a deep neural network-based polygraph scoring algorithm in Korea that outperformed both PolyScore and OSS-3.

Deep learning model development using Korean polygraph data. Performance comparison against established algorithms.

Published in 2025 in PMC/Sensors (2025), this work by Korean Research Team and colleagues examined this topic area. Key findings include: Developed a Korean polygraph scoring system using deep neural networks. DNN-based algorithm outperformed both PolyScore and OSS-3 on test data by accounting for bio-signal nonlinearity. This work contributed to the ongoing improvement of polygraph scoring methods, a critical factor in testing accuracy and reliability.

Deep learning approaches may represent the next generation of polygraph scoring, capable of extracting diagnostic information from nonlinear physiological patterns that elude conventional algorithms.

Background & Context

The objectivity and accuracy of polygraph testing have long depended on the skill of individual examiners who manually interpret physiological responses to identify deception. This subjectivity introduces vulnerabilities: examiner bias shaped by political, regional, or religious perspectives, as well as personal factors like fatigue and stress, can compromise judgment accuracy. Computerized scoring systems (CSSs) emerged as a solution to these problems, designed to automatically analyze polygraph charts and reduce human error. However, conventional CSS models that rely on linear classifiers struggle with the nonlinear nature of biological signals, resulting in poor performance. Systems like PolyScore and the Computerized Polygraph System (CPS) utilize linear logistic regression and linear discriminant analysis respectively, which excel at computational simplicity but falter when confronted with the complex, nonlinear dynamics inherent in autonomic physiological responses. This study addresses a critical gap in the field: the need to incorporate deep learning structures such as deep neural networks, which account for the nonlinearity of bio-signals, to enhance the effectiveness of computerized scoring systems. The research emerged from a broader initiative to develop a Korean polygraph system optimized for Korea's investigative environment, addressing limitations of imported systems that may not align with local testing protocols and population characteristics. The timing is particularly significant as artificial intelligence and deep learning have demonstrated transformative capabilities across medical diagnostics and sensor-based pattern recognition, yet their application to polygraph scoring has remained largely unexplored until now.

Research Design & Methodology

The study recruited physically and mentally healthy adults who volunteered for the experiment via university announcements and online postings. The experiment involved 91 participants, with 78 datasets ultimately used for analysis after excluding 13 datasets due to strong motion artifacts. Participants were classified into deception and non-deception groups, with their bio-signals recorded during a polygraph test conducted after engaging in simulated scenarios using the Utah technique, which is known for its high accuracy in Korea. The study received ethical approval from both the Institutional Review Board of the National Forensic Service and Wonju College, Yonsei University, with all participants providing informed written consent. The experimental protocol involved a mock theft scenario where participants either stole or did not steal a ring, then underwent polygraph testing denying knowledge of the theft. Each test comprised 10-12 questions spanning a total duration of 5 minutes, with only questions pre-reviewed during the pre-test interview allowed and participants restricted to "yes" or "no" responses. The researchers developed a custom polygraph device that measured five physiological signals: photoplethysmogram (PPG), blood pressure, thoracic respiration, abdominal respiration, and skin conductance. Notably, the inclusion of PPG alongside blood pressure represented an innovation, as conventional polygraphs typically measure only blood pressure for cardiovascular activity. The deep neural network architecture employed a sophisticated hybrid structure. The algorithm employed convolutional neural networks (CNNs) and recurrent neural networks (RNNs), with input data initially compressed through a CNN structure, after which the compressed time-series data were further processed using an RNN structure. Specifically, the system integrated five parallel four-layer CNNs (one for each physiological signal channel) with a Long Short-Term Memory (LSTM) unit—a specialized form of RNN particularly effective at capturing temporal dependencies in sequential data. Algorithm performance was assessed through recall, precision, and accuracy metrics using 10-fold cross-validation on data from 78 participants, which included 42 deceptive and 36 non-deceptive series, totaling 702 data points.

Results & Key Findings

The deep learning-based algorithm demonstrated exceptional performance that substantially exceeded conventional computerized scoring systems. The evaluation of the DNN structure incorporating five four-layer CNNs and one LSTM unit yielded a recall of 0.9681 ± 0.0314, precision of 0.9700 ± 0.0321, and F1 score of 0.9683 ± 0.0171. These metrics indicate the algorithm correctly identified approximately 97% of both deceptive and truthful examinees with remarkable consistency. The inclusion of LSTM proved essential: the DNN structure with only five four-layer CNNs (without LSTM) showed a recall of 0.9314 ± 0.0473, precision of 0.9377 ± 0.0529, and F1 score of 0.9340 ± 0.0449, underscoring the critical role of LSTM in enhancing performance. The comparative analysis against established algorithms revealed striking differences in real-world classification performance. In tests comprising 42 deception series and 36 non-deception series, PolyScore discriminated 17 deception series correctly but misclassified 6 and labeled 19 as inconclusive; it correctly identified 15 non-deception series, misclassified 5 as deceptive, and deemed 16 inconclusive. This translates to PolyScore producing inconclusive results in approximately 45% of cases—a substantial limitation in operational settings. OSS-3 discriminated 24 deception series, misclassified 18, and had no inconclusive outcomes, while correctly identifying 28 non-deception series, misclassifying 7 as deception, and marking 1 as inconclusive. OSS-3's lower inconclusive rate came at the cost of significantly higher error rates, with approximately 32% of cases misclassified overall. The photoplethysmogram signal proved surprisingly valuable. The developed algorithm excluding the PPG signal discriminated 36 out of 42 deception series with none misclassified and 5 inconclusive, and correctly identified 35 out of 36 non-deception series with none misclassified and 1 inconclusive. With the inclusion of the PPG signal, the algorithm discriminated 41 out of 42 deception series with none misclassified and only 1 inconclusive, and correctly identified all non-deception series. This represents near-perfect performance with only 1.3% inconclusive results and zero false positives or false negatives in the final configuration.

Discussion & Significance

This research demonstrates that deep learning architectures can extract diagnostic information from physiological signals that linear classifiers fundamentally cannot access. The performance gap between the DNN-based system and conventional algorithms is not merely incremental—it represents a qualitative leap in scoring capability. Where PolyScore produced inconclusive results in nearly half of examinations and OSS-3 misclassified roughly one-third of cases, the deep learning system achieved 98.7% definitive accuracy with minimal inconclusives and zero classification errors in its optimal configuration. This distinction matters enormously in operational contexts where inconclusive results necessitate retesting (consuming time and resources) and errors carry serious consequences for investigative and legal proceedings. The architectural choices reflect sophisticated understanding of physiological signal processing. The CNN layers effectively extract spatial features and local patterns within each physiological channel, compressing high-dimensional raw data into meaningful representations. The LSTM component then captures temporal dependencies and sequential patterns across the question series—critical for polygraph analysis where the relative magnitude and timing of responses to relevant versus comparison questions determines the outcome. The nonlinearity of bio-signals can be accommodated by integrating deep learning structures such as DNNs, allowing the system to model complex relationships between physiological arousal patterns and deceptive states that linear functions cannot represent. The inclusion of PPG alongside traditional cardiovascular measurement represents an important methodological innovation. While conventional polygraphs measure blood pressure via cuff, PPG uses optical sensors to detect blood volume changes in peripheral tissue. These signals capture complementary aspects of autonomic cardiovascular regulation, and the results demonstrate that this additional information stream meaningfully enhances classification accuracy. This finding suggests that expanding the physiological measurement space—when paired with algorithms capable of integrating multichannel data—can push polygraph accuracy toward theoretical limits.

Limitations & Considerations

The study had several limitations, particularly that to enhance and stabilize these scoring algorithms, the training dataset should be expanded. With 78 cases (42 deceptive, 36 non-deceptive) divided across 10-fold cross-validation, each training iteration utilized approximately 70 cases—a relatively modest sample for deep learning applications. While the cross-validation approach and consistent performance metrics suggest the model generalized well within this dataset, larger and more diverse training samples would strengthen confidence in real-world deployment. The participants were volunteers from university settings who engaged in mock theft scenarios with no genuine consequences, which may not fully replicate the psychological and physiological states of actual suspects in criminal investigations where stakes are considerably higher. The study compared the DNN system against PolyScore and OSS-3 using the same dataset from the Korean polygraph device measuring five channels including PPG. Since PolyScore and OSS-3 were designed for traditional three- or four-channel polygraphs, this comparison may not represent their optimal performance. Additionally, both conventional algorithms were developed and validated on Western populations, and their performance on Korean participants using Korea-specific testing protocols (the Utah technique as implemented locally) may differ from their typical accuracy ranges. The generalizability to other populations, testing techniques, and examination contexts requires further validation before the system could be adopted internationally.

Practical Applications

The immediate practical implication is that deep learning-based scoring can dramatically reduce inconclusive results while maintaining or improving accuracy compared to established systems. In operational polygraph programs—whether for criminal investigations, pre-employment screening, or post-conviction monitoring—the 45% inconclusive rate observed with PolyScore in this study would create significant logistical burdens and examination backlogs. A system that reduces inconclusives to under 2% while eliminating classification errors could transform operational efficiency and reduce the need for costly retesting. Integrating the developed algorithm directly into the polygraph software could support examiners with real-time evaluations, providing immediate feedback during testing sessions. For polygraph examiners and agencies considering adoption of such technology, several pathways forward emerge from this research. Connecting the developed polygraph to a central server via Ethernet and distributing it across investigation agencies can significantly increase data collection and training data volume, with future research focused on constructing a Korean computerized scoring network by connecting devices to a central server. This networked approach would enable continuous learning as the system encounters more cases, potentially adapting to regional variations and refining performance over time. For consumers of polygraph services, the research suggests that examinations utilizing validated deep learning scoring algorithms may offer higher accuracy and lower inconclusive rates than traditional manual scoring or older computerized systems—though the examiner's expertise in test administration, question formulation, and countermeasure detection remains irreplaceable regardless of scoring technology.