Introduction
Screening programs for transmissible diseases have become a major component of any coherent national and global health strategy. Elimination of viral hepatitis C (HCV) as a major public health threat by 2030 has become one of the main targets assigned by the World Health Organization in their long-term strategy, along with universal access to affordable and effective prevention, care, and treatment services [1]. The use of computer-aided and artificial intelligence in health-related resource management will greatly influence the chance of attaining these goals.
Aims & Methods
Our aim was to comparatively test multiple different design and training approaches for computer-assisted tools used in selecting at-risk populations during a large-scale regional screening for HCV infection. Between September 2020 and January 2023, we prospectively included participants over 40-years of age from the Southwestern region of Oltenia, Romania. All provided written informed consent and were presented with a 12-item rapid questionnaire followed by on-site anti-HCV antibody rapid testing. For the purposes of this study, the anonymized questionnaire answers, age, gender, and provenance of the subjects were collected, along with the result of the antibody testing results. We developed two artificial neural networks (ANN) models in Keras 2.7.0 with Tensorflow (Google, USA) using Python 3.7 (P.S.F., USA). One ANN model had access only to questionnaire answers (ANN1) and the other also included demographic data as inputs (ANN2). We trained each model on two sets of data – one with 15,139 participants (ANN1.1 and 2.1), the other with an additional 14,042 subjects (ANN2.1 and 2.2). All four resulting ANNs were tested on the same 6,226 additional participants’ datasets, comparing their accuracy, sensitivity, specificity, negative and predictive values at five confidence intervals – 0.6, 0.7, 0.8 and 0.9, also accounting for the number of tests needed by the systems each time to find true positives.
Results
The training datasets contained 15,139 and 29,181 subjects, respectively. The total screened population was 35,407, including 6,226 in the test lot (42 HCV positives). Table 1 shows the number of required tests by each ANN model, and the number of true positives. The ANN not accounting for demographic data performed better in predicting a higher number of positives, at lower confidence levels and with a high usage of tests, while the complete model consistently required less tests, with less true positives at lower confidence intervals, but significantly better results at higher confidence intervals. Overall, the most balanced results were obtained by ANN2.2 (all inputs and data) – 69% sensitivity, 56.4% specificity, 1% PPV, 99.6% NPV and 56.5% accuracy at 0.6 and 26.2%, 97.4%, 4.2%, 99.4% and 96.9%, respectively, at 0.9 confidence level.
| ANN1.1 | ANN1.2 | ANN2.1 | ANN2.2 |
| Tests | Positives | Tests | Positives | Tests | Positives | Tests | Positives |
| 0.6 | 4486 | 38 | 3875 | 36 | 3675 | 28 | 2723 | 29 |
| 0.7 | 432 | 14 | 528 | 14 | 2133 | 23 | 1651 | 24 |
| 0.8 | 53 | 1 | 446 | 14 | 274 | 6 | 352 | 11 |
| 0.9 | 5 | 0 | 28 | 1 | 80 | 0 | 165 | 7 |
Conclusion
The ANN system trained on a larger dataset and accounting for more variables proved more valuable in a low-resource environment, while a limited system can be more broadly applied in smaller programs, at higher costs but with more positives found. The use of such tools to select at-risk populations in HCV screenings can facilitate fast, reliable detection and access of a larger population in a resource-constrained situation or special environment.
References
https://apps.who.int/iris/bitstream/handle/10665/246177/WHO-HIV-2016.06-eng.pdf
