Abstract—Cancer is one of the most common and serious medical conditions with more than 144 000 Australians having been diagnosed with cancer in 2019. The non-specific nature of cancer symptoms and its low prevalence make cancer diagnosis particularly challenging, especially for primary care physicians/General Practitioners (GPs). Ongoing research in cancer diagnosis places a heavy focus on understanding the epidemiology of cancer symptoms. With GPs being the first point of contact for most patients, prediction models using the patient’s medical history from primary care data can be a useful decision tool for early cancer detection. Our work both investigates the opportunities to use primary care data, specifically pathology data, for developing such decision tools and tackles the challenges coming from uncertainty in the data such as irregular pathology records. We present opportunities using the results within the frequently ordered full blood count to determine relevance to a future cancer diagnosis. By using several different pathology metrics, we show how we can generate features suitable for AI models that can be used to detect cancer 3 months earlier than current practices. Though the work focuses on patients with lung cancer, the methodology can be adjusted to other types of cancer and other data within the medical records. Our findings demonstrate that even when working with incomplete or obscure patient history, hematological measures contain valuable information that can indicate the potential of cancer diagnosis for up to 8 out of 10 patients. The use of the proposed decision tool presents a way to incorporate pathology data in the current cancer diagnosis practices and to incorporate various pathology tests or other primary care datasets for similar purposes.
 

Index Terms—explainable AI, early cancer detection, uncertainty in data, feature generation