Estimating national Red List statuses for fungi in Sweden - An improved deep learning approach to account for unbalanced data

National and international extinction threat assessments provide critical data for research and conservation. Yet, many taxa remain unassessed due to the intensive data requirements of manual evaluations. For example, in Sweden, fewer than 40 % of known fungal species have an assigned extinction risk in the national Red List. To address this, approximation methods have emerged to provide estimates of extinction risk for species lacking an officially assigned status. These automated extinction threat inferences can be viewed as a best-evidence approximation, flagging potentially threatened species for further ecological research and targeted manual Red List assessments. One such approximation method is the deep learning approach implemented in the IUCNN package, predicting extinction risk using environment features extracted from species occurrence data. However, such approximation models face challenges arising from class imbalances, with most species classified as Least Concern, which biases predictions. Here, we advance the IUCNN framework by applying balanced accuracy as a performance metric and using class weights and oversampling to mitigate class imbalance. This approach improves the model's ability to predict intermediate classes, such as Near Threatened (NT), Vulnerable (VU), and Endangered (EN), resulting in higher overall balanced accuracy. We applied this approach nationally to 1445 unassessed Swedish fungal species, covering nearly half of the country's known fungi and increasing the number of fungi with a national extinction threat status by over 10 %. Our approach identified 661 species as potentially threatened, doubling the total of fungal species listed as threatened compared to the existing Swedish Red List assessments. This highlights previously unrecognized conservation priorities by shortlisting potentially threatened species that would be good targets to receive more attention from conservation authorities and manual Red List assessments. • First implementation of national-level Red List status predictions • Improved performance of IUCNN model for minority-classes by accounting for imbalanced data • National Red List status predicted for 1445 previously unassessed fungal species • Applying high spatial resolution prediction features for national application