Deep Learning Model Meets Community-based Surveillance of Acute Flaccid Paralysis

"[T]his work represents a significant step toward leveraging artificial intelligence for community-based AFP [acute flaccid paralysis] surveillance from images, with substantial implications for addressing global health challenges and disease eradication strategies."

Acute flaccid paralysis (AFP) surveillance involves monitoring and reporting cases of AFP, which could be indicative of poliovirus infection. The surveillance system typically involves health workers and surveillance officers at various levels of the healthcare system, including hospitals, health centres, and communities. The community-based surveillance system implemented in Ethiopia has significantly improved AFP surveillance. However, challenges like delayed detection and disorganised communication persist. This work proposes a simple deep learning model for AFP surveillance, leveraging transfer learning on images collected from Ethiopia's community key informants through mobile phones.

The rationale for using artificial intelligence (AI) for AFP surveillance, as this study proposes, is that AFP case images are rare because:
 

• Polio is eradicated from many countries except a few endemic countries; as a result, it is difficult to collect images of positive AFP cases.
• There are ethical issues related to personal information uses.
• The cases are available in low-income countries, where it is expensive to collect and store AFP case images.
• Long-term investment is needed to have a surveillance system that works in coordination to collect and store the images.

That is, it is hard to find large number of AFP case images that can be used to train the data that intensive deep learning models need in order to extract important features and generalise well when subjected to new, previously unseen images of AFP. Thus, this work sought to develop a simple transfer learning for the early detection of AFP-suspected cases from images collected by community volunteers (CVs) using vision transformer architectures.

The data for this study were collected from the CORE Group Partner Project (CGPP) Ethiopia implementation catchment area. The CGPP Ethiopia was established in 1999 and started implementation in Ethiopia in November 2001. CGPP Ethiopia has supported and coordinated efforts of private voluntary organisations (PVOs) and non-governmental organisations (NGOs) involved in polio eradication activities. CGPP trained and dispatched more than 10,000 CVs to be able to immediately report AFP cases in their communities. The CVs carry out community-based surveillance by searching house to house for children with symptoms of AFP. CGPP Ethiopia deployed more than 10,000 CVs in more than 1,700 Kebeles (the grassroot administration body in Ethiopia). For this study, an AFP surveillance dataset of 428 images, comprising 228 suspected AFP cases and 200 normal cases, was created using images of Ethiopian children collected over the past five years through the CGPP Ethiopia's community-based surveillance system. The images were anonymised by removing the parts above neck, and no additional personal information was used for this study.

This study utilised the pretrained vision transformer model for distinguishing between normal and suspected AFP images. The transfer learning approach was implemented using a vision transformer model pretrained on the ImageNet dataset. The study found that the proposed model outperformed convolutional neural network-based deep learning models and vision transformer models trained from scratch, achieving superior accuracy, F1-score, precision, recall, and area under the receiver operating characteristic curve (AUC). It emerged as the optimal model, demonstrating the highest average AUC of 0.870 ± 0.01. Statistical analysis confirmed the significant superiority of the proposed model over alternative approaches (P < 0.001).

In short, by bridging community reporting with health system response, this study hopes to offer a scalable solution for enhancing AFP surveillance in low-resource settings. The study is limited in terms of the quality of image data collected, necessitating future work on improving data quality. The establishment of a dedicated platform that facilitates data storage, analysis, and future learning can strengthen data quality.

The researchers conclude by arguing: "Overall, this work has the potential to significantly improve AFP surveillance and contribute to the broader efforts in global health security and disease eradication."

Full list of authors, with institutional affiliations: Gelan Ayana, Jimma University, Africa-Canada Artificial Intelligence and Data Innovation Consortium (ACADIC), and Global South Artificial Intelligence for Pandemic and Epidemic Preparedness and Response Network (AI4PEP); Kokeb Dese, Jimma University, ACADIC, and AI4PEP; Hundessa Daba Nemomssa, Jimma University, ACADIC, and AI4PEP; Hamdia Murad, Jimma University, ACADIC, and AI4PEP; Efrem Wakjira, Jimma University, ACADIC, and AI4PEP; Gashaw Demlew, Jimma University, ACADIC, and AI4PEP; Dessalew Yohannes, Jimma University, ACADIC, and AI4PEP; Ketema Lemma Abdi, Jimma University, ACADIC, and AI4PEP; Elbetel Taye, Ethiopian Artificial Intelligence Institute, ACADIC, and AI4PEP; Filimona Bisrat, CORE Group Partner Project; Tenager Tadesse, CORE Group Partner Project; Legesse Kidanne, CORE Group Partner Project; Se-woon Choe, Kumoh National Institute of Technology; Netsanet Workneh Gidi, Jimma University; Bontu Habtamu, ACADIC and AI4PEP; and Jude Kong, University of Toronto, ACADIC, and AI4PEP