Prediction of lymph node parasite load from clinical data in dogs with leishmaniasis: An application of radial basis artificial neural networks

• Published in: Veterinary parasitology 2017-01, Vol.234, p.13-18
• Main Authors: Torrecilha, Rafaela Beatriz Pintor, Utsunomiya, Yuri Tani, Batista, Luís Fábio da Silva, Bosco, Anelise Maria, Nunes, Cáris Maroni, Ciarlini, Paulo César, Laurenti, Márcia Dalastra
• Format: Article
• Language: English
• Subjects: Animals; Canis lupus familiaris; Dog Diseases - diagnosis; Dog Diseases - parasitology; Dog Diseases - pathology; Dogs; Leishmania spp; Leishmaniasis - diagnosis; Leishmaniasis - parasitology; Leishmaniasis - pathology; Leishmaniasis - veterinary; Lymph Nodes - parasitology; Machine learning; Neural Networks (Computer); Parasite Load - veterinary; qPCR
• ISSN: 0304-4017; 1873-2550
• DOI: 10.1016/j.vetpar.2016.12.016

•Leishmania infantum load was predicted from clinical data by a radial basis artificial neural network.•Predictions were unbiased and non-inflated, with an accuracy of 0.869.•When qPCR is unavailable, parasite load can be estimated using machine learning techniques. Quantification of Leishmania infantum load via real-time quantitative polymerase chain reaction (qPCR) in lymph node aspirates is an accurate tool for diagnostics, surveillance and therapeutics follow-up in dogs with leishmaniasis. However, qPCR requires infrastructure and technical training that is not always available commercially or in public services. Here, we used a machine learning technique, namely Radial Basis Artificial Neural Network, to assess whether parasite load could be learned from clinical data (serological test, biochemical markers and physical signs). By comparing 18 different combinations of input clinical data, we found that parasite load can be accurately predicted using a relatively small reference set of 35 naturally infected dogs and 20 controls. In the best case scenario (use of all clinical data), predictions presented no bias or inflation and an accuracy (i.e., correlation between true and predicted values) of 0.869, corresponding to an average error of ±38.2 parasites per unit of volume. We conclude that reasonable estimates of L. infantum load from lymph node aspirates can be obtained from clinical records when qPCR services are not available.