Use of deep learning to predict chronic wasting disease status based on animal movement Authors: Raymond L. Blaha, Christopher J. Silva, Stephanie A. Cunningham, Melia T. DeVivo, David R. Edmunds, Melanie R. Boudreau Published: 2026 ABSTRACT Chronic Wasting Disease (CWD) is an invariably fatal prion disease that impacts cervid populations and wildlife management across North America. Infected cervids often remain asymptomatic for months and movement-based anomaly detection from Global Positioning System collaring data offers a potential tool for understanding early and late stage CWD-based behavioral changes. Here we evaluate whether deep learning behavioral anomaly detection models such as autoencoders (AE) and conditional autoencoders (cAE) can effectively identify anomalous movement changes in free-ranging mule deer (Odocoileus hemionus) that may have been associated with CWD infection. Unsupervised AEs achieved ≥ 84% accuracy and 89% precision irrespective of whether the model was trained using only CWD− or a 60/40 split of CWD+ / CWD− individuals. Important movement features in distinguishing between CWD+ and CWD− animals included metrics related to velocity, direction, and sinuosity. In contrast, supervised cAEs only achieved 54–77% accuracy and ≤ 53% precision across models trained with incidence rates of CWD+ individuals ranging from 10 to 40%; the models also had inconsistent results reducing their generality. Finally, using the best fitting model (AE trained using only CWD− animals), we found that early-stage animals exhibited reduced space use, whereas late-stage individuals presented more pronounced declines in velocity and altered directional patterns. These findings indicate that AEs can accurately identify CWD-related behavioral anomalies using movement data and that variation in what movement metrics matter depends on the stage of the disease.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40462-026-00668-4. Background Movement ecology provides a framework for understanding how host behavior influences disease dynamics in free ranging wildlife. Recent work has emphasized the role of movement processes in shaping spatio-temporal patterns of transmission, contact structure, and exposure risk, highlighting the value of integrating movement analyses into disease research [1]. For some diseases, infection can alter host movement behavior such that exposure and transmission rates are impacted by both direct and indirect transmission pathways [1]. These movement-mediated processes can scale to influence disease dynamics at the landscape level, and understanding these dynamics can improve our ability to model and manage disease spread [2]. However, a key challenge remains in translating complex movement data into reliable indicators of disease presence, particularly when behavioral changes are subtle, heterogeneous, or emerge gradually over time. Chronic wasting disease (CWD) is an invariably fatal, transmissible prion disease affecting cervids [3]. In the United States, CWD prevalence in wild populations is dependent on location and can vary as low as < 1% and as high as 60% [4, 5]. High infection rates have significant population level consequences. For instance, CWD has been found to be the primary driver of long-term population decline in a Wyoming mule deer (Odocoileus hemonius) herd, reducing population growth by 20–25% annually [6] and population models indicate declines can occur when CWD prevalence reaches ~ 25% [7]. As CWD spreads, it not only impacts the infected individuals and populations, but also federal and state agencies trying to manage those populations, as well as public consumptive and non-consumptive wildlife users [8–13]. Infected cervids may not show obvious signs of CWD infection for many months despite being infectious, making monitoring the spread of CWD challenging [9, 14]. Naturally infected deer succumb to CWD between 1 and 20 months (average 12 ± 5 months) after testing positive for CWD, although exact timing can be variable depending on the level of prion exposure and the stage of CWD at which the deer tested positive [15]. There are indications that animals in the preclinical phase of the disease show subtle behavioral changes, including altered interactions with handlers and other animals, and reduced appetite (as seen in mule deer and elk [Cervus canadensis]; [16, 17]). In advanced stages of the disease clinical signs are more prominent and include impaired motor function (such as flaccid face muscles, excessive saliva production, and grinding of teeth), changes to spatial behavior (e.g., apparent unawareness, prolonged standing with lowered head, walking in a particular pattern, reduced social interactions, and reduced space use), altered activity rhythms (e.g., extended periods of inactivity or changes to circadian rhythm), altered habitat use, and progressive weight loss that inevitably ends with the animal’s death [3, 18–21]. Deer in the later stages of the disease are also > 1.6 times more susceptible to deer-vehicle collisions [22] and CWD-infected deer are ~ 4 times more susceptible to mountain lion predation [23, 24] and 1.7 times more likely to be harvested by hunters [9] than uninfected deer. However, by the time these advanced clinical signs become observable, or have an impact on the lives of infected cervids, they have been shedding CWD prions into the environment for months. Developing methods to identify nuanced behavioral anomalies associated with early stages of CWD remains a unique challenge, particularly because data from infected deer are severely limited given agency requirements to remove known infected individuals off the landscape. Consequently, datasets that contain longitudinal, high resolution movement data, and confirmed CWD status are exceedingly rare and uniquely valuable for investigating preclinical signs. Even when available, behavioral data are limited to movements of animals (i.e., Global Positioning System (GPS) collar data that are not supplemented with other biologging information) and work to date has only focused on the drastic movement changes during clinical signs of CWD, which are easier to detect using traditional ecological analyses [20]. Finally, advanced machine learning methods can be used to tease out subtle behavioral patterns, but they often require sufficiently large sets of labeled data (i.e., known CWD status) for model training and struggle with high-dimensional information [25, 26]. Autoencoders (AEs) are deep learning models that can be used to understand complex patterns in high-dimensional data. As such, they may help to identify more subtle changes in behavior before the clinical neurological stage, which has not been possible using other methods. Using GPS-based movement data, we evaluated the efficacy of autoencoders [27] in identifying behavioral differences between CWD-positive (CWD+) and CWD-negative (CWD−) individuals. Autoencoders are deep learning models that can extract latent representations that characterize “normal” behavior, making them especially well-suited for detecting anomalies from complex, high-dimensional metrics created from movement datasets. In addition, AEs offer the flexibility to be trained in either an unsupervised or semi-supervised manner, therefore not requiring labels (e.g., CWD status), but instead learning from dominant patterns present in the training set. However, if labels are provided, conditional autoencoders (cAE) can leverage this information to potentially improve detection of behavioral anomalies [27]. As a case study, we used a multi-year dataset collected from mule deer in southcentral Wyoming that included both movement histories and confirmed infection outcomes, offering a rare opportunity to examine behavioral differences associated with disease progression [6]. Specifically, our objectives were to determine whether (1) AE models could quantify movement patterns deviating from established CWD− behavioral baselines, (2) we could use deviations to accurately distinguish between CWD+ and CWD− individuals, and (3) deviations could also be used to examine behavioral differences between early- and late-stage CWD+ deer. We compared AE model performance to typical modelling approaches used to examine binary responses in animals to ensure that an AE modelling approach was necessary. We expected that individuals whose movement deviated substantially from a learned baseline would produce consistent behavioral anomalies and that we could use this information in subsequent identification of disease-induced behavioral shifts. We also expected that relationships between disease status and movement would be complex, meaning that the AE would outperform conventional approaches and warrant implementation. Methods Study area and focal population The data used in this study were obtained from a previously published study [6]. Briefly, 118 females and 25 males from the South Converse Mule Deer Herd in southcentral Wyoming were captured via aerial net-gunning in February and March of each year (2011–2014), chemically immobilized, and fitted with GPS collars (Lotek Wireless Inc., Newmarket, Ontario, Canada) or very high frequency (VHF) radio-collars (Advanced Telemetry Systems, Inc., Isanti, Minnesota, USA). Collars were programmed to collect fixes at 6-hour intervals and contained a mortality signal that was activated after 4 h of inactivity. Deer were tested for CWD using tonsillar biopsies and immunohistochemistry staining for infectious prion proteins [14] during the initial collaring and during their annual recapture. When possible, postmortem confirmation of CWD status was done using immunohistochemistry and an enzyme-linked immunosorbent assay of tissue samples from retropharyngeal lymph nodes, tonsils, and the obex region of the medulla oblongata [6]. Further details regarding study area, CWD sampling, and mule deer capture and handling are provided in [6]. We note that CWD− and CWD+ individuals were distributed evenly across the study area. Data cleaning and feature engineering for movement metrics We first filtered GPS tracking data to ensure accuracy and consistency prior to movement modeling, removing all fixes with dilution of precision > 10 [28]. We filtered the dataset to retain only individuals with GPS data spanning at least 20 days, with each day contributing at least 3 of the 4 expected location fixes (n = 107). This threshold aimed to balance data quality with sample retention, while minimizing noise from animals with short or highly fragmented tracking histories. We then constructed movement trajectories for each deer using the adehabitatLT package v. 0.3.28 and regularized trajectories to ensure a consistent fix rate [29]. As mule deer sometimes migrate between summer and winter ranges [30], we manually classified each mule deer as being in a resident or migratory state by visually inspecting net squared displacement (calculated from trajectories) across each individual’s entire location history. To not confound our analyses with movements that may reflect characteristics of a given state (e.g., more tortuous movements during residency vs. directed movements during active migration), we removed locations taken during migration, leaving only information from time spent on winter and summer ranges. For individuals that were residential throughout the year, we split the year into two ranges (winter and summer) to account for ecological variation in movement behavior across the year using the median seasonal start and end dates from migratory deer (i.e., summer: 21-Apr to 25-Sep, winter: 26-Sep to 20-Apr). For each deer, we also excluded portions of their trajectories that were part of short-term exploratory movements (i.e., excursions or random walks) and coded data to exclude these movements from calculations. To capture fine-scale movement trends, we calculated a suite of step-level and window-based features for each deer. For each location, we derived multiple stepwise movement metrics, including velocity, relative turning angle, and sinuosity (Table 1). We then captured movement trends occurring at a longer time scale by calculating 7-day moving window averages for velocity and turn angle to quantify rolling means, standard deviations, skewness, and kurtosis. We also quantified path sinuosity over each window and, finally, weekly 95% autocorrelated kernel density estimates (AKDE) using the ctmm package v. 1.2.0 [31]. For each AKDE, we quantified the estimated area and the 95% confidence interval of that area estimate. Ultimately, we obtained a set of 16 metrics that we used as input for our models. All metrics were expected to reflect not only shifts in overall activity, but also increased movement inconsistencies, or behavioral disruptions possibly associated with CWD. All data cleaning and movement metric calculations were performed in R v. 4.4.2 [32], additionally using packages sf v. 1.0.19 [33, 34]., tidyverse v. 2.0.0 and ggplot2 v. 3.5.1 [35]. We also used sklearn 1.6.1 MinMax scaler to standardize the data from [0,1] in Python v. 3.11.11 before use in the AE [36]. Table 1Definitions of various movement metrics and how they were calculatedMetricDefinitionFormula / MethodologyAbsolute Value of VelocityInstantaneous movement speed between sequential GPS fixes\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{t}=\:\frac{{d}_{t}}{{\varDelta\:t}_{t}}\:,\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{t}$$\end{document} is the Euclidean distance and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varDelta\:t}_{t}$$\end{document} is time between t− 1 and t.Absolute Value of Relative Turn AngleAbsolute change in bearing between consecutive movement vectors\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Theta\:}_{t}=|\left(\left({\alpha\:}_{t}-\:{\alpha\:}_{t-1}+180\right)\:mod\:360\right)-180|$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{t}=atan2({y}_{t}-{y}_{t-1},\:\:{x}_{t}-{x}_{t-1})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{t-1}={\alpha\:}_{t}=atan2\left({y}_{t-1}-{y}_{t-2},\:\:{x}_{t-1}-{x}_{t-2}\right)\:$$\end{document} are successive movement bearings and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}\in\:[0,\:180]$$\end{document}.SinuosityA measure of movement straightness calculated as the net displacement divided by cumulative path length over prior 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{v,t}=\:\frac{\sqrt{{({x}_{t}-\:{x}_{t-28})}^{2}+\:{({y}_{t}-\:{y}_{t-28})}^{2}}}{\sum\:_{t28}^{t}{d}_{i}}$$\end{document}, where x and y are planer coordinates and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}$$\end{document} is step length at time t.Rolling Velocity or Relative Turning Angle MeanRolling mean of velocity or relative turn angle over 28 observations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}{v}_{t}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Theta}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28\:}^{t-1}{\Theta\:}_{t},\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle at time i.Rolling Velocity or Relative Turning Angle Standard DeviationRolling standard deviation of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{v,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(v}_{i}-\:{\mu\:}_{v,t}{)}^{2}}$$\end{document}or\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{\Theta\:,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(\Theta\:}_{i}-\:{\mu\:}_{\Theta\:,t}{)}^{2}\:}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} denotes the corresponding rolling mean.Rolling Velocity or Relative Turning Angle SkewnessRolling skewness of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Skew}_{z,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}\:\frac{{(z}_{i}-\:{\mu\:}_{z,t}{)}^{3}}{\:{{\sigma\:}^{3}}_{z,t}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{i}$$\end{document} denotes either velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document}) or absolute relative turn angle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document}), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{z,t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{z,t}$$\end{document} are the corresponding rolling mean and standard deviation.Rolling Velocity or Relative Turning Angle KurtosisRolling kurtosis of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Kurtosis}_{z,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}\:\frac{{(z}_{i}-\:{\mu\:}_{z,t}{)}^{4}}{\:{{\sigma\:}^{4}}_{z,t}}-3\:$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{z}_{i}$$\end{document} denotes either velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document}) or absolute relative turn angle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document}), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{z,t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{z,t}$$\end{document} are the corresponding rolling mean and standard deviation.AKDE Area Estimates (Low, Estimate, High)95th confidence interval and point estimate of utilization distribution over 7-day windowsEstimated using AKDE (Autocorrelated Kernal Density Estimation) per the ctmm R package Table 1Definitions of various movement metrics and how they were calculatedMetricDefinitionFormula / MethodologyAbsolute Value of VelocityInstantaneous movement speed between sequential GPS fixes\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{t}=\:\frac{{d}_{t}}{{\varDelta\:t}_{t}}\:,\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{t}$$\end{document} is the Euclidean distance and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varDelta\:t}_{t}$$\end{document} is time between t− 1 and t.Absolute Value of Relative Turn AngleAbsolute change in bearing between consecutive movement vectors\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Theta\:}_{t}=|\left(\left({\alpha\:}_{t}-\:{\alpha\:}_{t-1}+180\right)\:mod\:360\right)-180|$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{t}=atan2({y}_{t}-{y}_{t-1},\:\:{x}_{t}-{x}_{t-1})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{t-1}={\alpha\:}_{t}=atan2\left({y}_{t-1}-{y}_{t-2},\:\:{x}_{t-1}-{x}_{t-2}\right)\:$$\end{document} are successive movement bearings and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}\in\:[0,\:180]$$\end{document}.SinuosityA measure of movement straightness calculated as the net displacement divided by cumulative path length over prior 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{v,t}=\:\frac{\sqrt{{({x}_{t}-\:{x}_{t-28})}^{2}+\:{({y}_{t}-\:{y}_{t-28})}^{2}}}{\sum\:_{t28}^{t}{d}_{i}}$$\end{document}, where x and y are planer coordinates and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}$$\end{document} is step length at time t.Rolling Velocity or Relative Turning Angle MeanRolling mean of velocity or relative turn angle over 28 observations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}{v}_{t}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Theta}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28\:}^{t-1}{\Theta\:}_{t},\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle at time i.Rolling Velocity or Relative Turning Angle Standard DeviationRolling standard deviation of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{v,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(v}_{i}-\:{\mu\:}_{v,t}{)}^{2}}$$\end{document}or\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{\Theta\:,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(\Theta\:}_{i}-\:{\mu\:}_{\Theta\:,t}{)}^{2}\:}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} denotes the corresponding rolling mean.Rolling Velocity or Relative Turning Angle SkewnessRolling skewness of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Skew}_{z,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}\:\frac{{(z}_{i}-\:{\mu\:}_{z,t}{)}^{3}}{\:{{\sigma\:}^{3}}_{z,t}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{i}$$\end{document} denotes either velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document}) or absolute relative turn angle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document}), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{z,t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{z,t}$$\end{document} are the corresponding rolling mean and standard deviation.Rolling Velocity or Relative Turning Angle KurtosisRolling kurtosis of velocity or relative turn angle over 28 observations\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Kurtosis}_{z,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}\:\frac{{(z}_{i}-\:{\mu\:}_{z,t}{)}^{4}}{\:{{\sigma\:}^{4}}_{z,t}}-3\:$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{z}_{i}$$\end{document} denotes either velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document}) or absolute relative turn angle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document}), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{z,t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{z,t}$$\end{document} are the corresponding rolling mean and standard deviation.AKDE Area Estimates (Low, Estimate, High)95th confidence interval and point estimate of utilization distribution over 7-day windowsEstimated using AKDE (Autocorrelated Kernal Density Estimation) per the ctmm R package Definitions of various movement metrics and how they were calculated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28}^{t-1}{v}_{t}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Theta}_{v,t}=\:\frac{1}{28}\sum\:_{i=t-28\:}^{t-1}{\Theta\:}_{t},\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle at time i. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{v,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(v}_{i}-\:{\mu\:}_{v,t}{)}^{2}}$$\end{document}or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{\Theta\:,t}=\:\sqrt{\frac{1}{28}\sum\:_{i=t-28}^{t-1}{(\Theta\:}_{i}-\:{\mu\:}_{\Theta\:,t}{)}^{2}\:}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} is velocity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} is the absolute relative turn angle, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} denotes the corresponding rolling mean. Baseline models As a means of evaluating whether a complex model such as the AE was warranted to distinguish between CWD− and CWD+ individuals, we implemented a logistic regression with disease status (CWD− deer = 0, CWD+ deer = 1) as the response variable and our movement metrics as the predictor variables. We also implemented a single feature boundary threshold classifier model with disease status (CWD− deer = 0, CWD+ deer = 1) as the response variable and each individual movement metric as the predictor variable. In both cases, we implemented a 80:20 split of the individuals in the data for use in model testing. Models were run using the caret R package [37]. Autoencoder architecture Autoencoders are a class of neural network models designed to learn compact representations of high-dimensional data [38–41]. An AE consists of two components: an encoder, which maps the input data to a lower-dimensional latent representation, and a decoder which reconstructs the input from this representation [42]. The input into the AE consists of the movement features derived from our GPS data (i.e., the 16 calculated movement metrics; Table 1) and we implemented and compared two types of autoencoders, an unsupervised AE and a supervised (or conditional) AE (hereafter referred to ‘cAE’; Figure S1). In an AE, anomaly detection begins when the AE is trained on data representing a baseline. Under this setting, the model learns to reconstruct typical patterns present in the training data. Observations that differ from this baseline may exhibit larger reconstruction errors; thus, reconstruction error can be used as a proxy measure for deviation from expected behavior [43]. Testing data is then used to assess model performance. We constructed two types of unsupervised AE, which were trained and tested in two different ways. In the first model, we used all CWD− deer for training and tested on CWD+ deer (i.e., a full holdout). In the second model, we trained the AE using 60% of all CWD− deer and tested on all CWD+ plus the remaining 40% of CWD− individuals (60/40 split). The AE was parameterized to have fully connected dense neural layers with rectified linear unit (ReLU) activations in both the encoder and decoder. It also had a custom loss function (Supplementary Information) that looks for deviations from the mean squared error and adds a penalty term for large reconstruction errors, making the model more sensitive to rare abnormal movement patterns that might not be present in the CWD− training set. AEs can be hyper-parameter tuned for an optimum number of neural layers which can prevent overfitting while also still having the ability to capture more granular differences in animal behavior associated with disease progression, and our AE was tuned to have 32-16-4-16-32 neural layers (Figure S1). Hyper-parameter tuning also allowed us to set the custom loss penalty term to 0.5, a learning rate of 0.0001, and batch size of 128. We implemented TensorFlow’s Early Stopping to prevent overfitting and extensive computational time using a patience of 5 epochs over 100 total epochs. Conditional autoencoder architecture Autoencoders can also be implemented with supervised training, and we implemented a cAE framework to incorporate disease stage information. The cAE models followed the same structure as the AE models, but they are augmented with an explicit predictor with the disease status label, enabling the model to learn specified CWD disease input (i.e., CWD− deer = 0, CWD+ deer = 1). The cAE was also parameterized to have fully connected dense layers with ReLU activations and a custom loss function which was hyper-parameter tuned to have a penalty loss of 0.5. It was also tuned to have 64-32-8-32-64 neural layers, a learning rate of 0.0001, and batch size of 128 (Figure S2). Because the cAE uses labeled data, we could simulate various levels of CWD prevalence in the model to understand how prevalence would influence model performance. We used the following range of prevalence levels in our training and testing data to reflect realistic disease prevalence scenarios: 10%, 20%, 30% or 40%, with 40% being one of the higher prevalences documented in the wild [6]. We implemented a stratified 3-fold for 10% prevalence experiment, 4-fold for 20% experiment, and 5-fold cross-validation for 30% and 40% experiments, with folds based on individual animal IDs and status labels. The training and testing datasets were constructed to match infection rates but included a similar number of IDs in each dataset (e.g., a 1:1 ratio; for the 10% training dataset having 3 infected to 28 non-infected in training and testing). At each fold, the model was trained and validated on disjoint sets, with class imbalance addressed via sample weighting. For the AE and cAE, model development and training were performed using TensorFlow version 2.19.0 [44] in Python v. 3.11.11. Thresholds for CWD status evaluation After training, we set decision rules prior to testing and evaluation by defining a decision policy to allow reconstruction errors to be used to redefine CWD− vs. CWD+ status for each individual. Based on visual assessment of the reconstruction errors, we noted that reconstruction errors that were in the top 25th percentile were related to rare outlier instances, therefore, to prevent rare, high error outliers from artificially inflating the decision threshold, we excluded the top 25% of reconstruction errors from this calculation. The anomaly threshold was then defined as the top 90th percentile of the remaining errors, and observations exceeding this value were classified as anomalous. However, an individual was only classified as CWD+ if data were flagged as anomalous and either the number of anomalous observations exceeded 75 fixes (> 25 days), or at least 75% of that individual’s observations were classified as anomalous. This ensured that we were not flagging anomalous behavior that may arise from individuality, but rather consistent changes in behavior over time (Figure S1). For the cAEs, each individuals’ data was reconstructed twice, once through the CWD− and once through the CWD+ pipeline. In theory, CWD+ deer should have a smaller error when the model is learning information from CWD+ individuals, and vice versa for CWD− deer. To quantify this difference, we defined a reconstruction margin as the difference between reconstruction error under the CWD− condition and reconstruction error under the CWD+ condition. Positive margins therefore indicated observations that were better reconstructed under the CWD+ condition corresponded to stronger evidence of CWD+ behavior (Figure S1). Margins were then averaged across observations for each individual to obtain an individual level score. Again, we noted that reconstruction errors that were in the top 27.5th percentile were related to rare outlier instances, therefore, to prevent rare, high error outliers from artificially inflating the decision threshold, we excluded the top 27.5th of reconstruction errors from this calculation. We hyper tuned these thresholds to lower the false positive rate for each of the cAEs and improve stability across folds. The final threshold was defined as the median across cross-validation folds. Again, an individual was only classified as CWD+ if either the number of anomalous observations exceeded 75 fixes (> 25 days), or at least 75% of that individual’s observations were classified as anomalous to ensure that we were examining consistent changes in behavior over time (Figure S1). For both the AE and cAE models, we then used the testing dataset to determine the accuracy and precision of the models using the reconstruction error thresholds determined from training and validation (Figure S2). Model evaluation For both the baseline and autoencoder models, we used the testing data to evaluate model performance by identifying the number of correctly classified CWD− deer (true negative; TN), the number of correctly identified CWD+ deer (true positive; TP), the number of missed CWD+ deer (false negatives; FN), and the number of CWD− deer flagged as CWD+ (false positive; FP). We used these values to obtain sensitivity (i.e., the true positive rate calculated as TP/(TP + FN)) and specificity (i.e., the true negative rate TN/(TN + FP)). Using these, we calculated model accuracy ((TP + TN)/(TP + TN + FP + FN)). For the AE and cAE models we also calculated precision (TP/(TP + FP)), along with the true skill statistic (TSS) using sensitivity and specificity (TSS = sensitivity + specificity − 1). While accuracy and precision values range from 0 to 100%, TSS values range from 1 to − 1, with a value of 1 indicating perfect agreement and values ≤ 0 being equivalent to random predictions. In addition, for the cAEs, we also report prevalence identified (#CWD+/(#CWD+ + CWD−)). After obtaining testing metrics for each model, we only examined movement metrics that had anomalous reconstructions errors that encompassed ≥ 10% of the change in behavior as potentially having some biological significance. We then inversely scaled the data from the best model to the original movement metric values to explore potential differences among three CWD status groups: (i) CWD− individuals throughout the study period, (ii) “early stage” CWD− individuals that tested positive at some point during annual monitoring, and (iii) “late stage” individuals that tested CWD+ at the time of first capture. Since the exact timing of initial infection is unknown in free ranging systems, these classifications do not correspond to precise durations post infection. However, since CWD progression typically occurs over approximately 12 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\pm\:\:$$\end{document}5 months following detectable infection [15], these groupings likely represent individuals in earlier versus later stages of disease progression. Results After data cleaning and filtering, we were able to use 89 female and 18 male mule deer in our analyses. The average annual CWD prevalence across these animals was 24% (18% females, 43% males). Among the 107 mule deer with known survival outcomes, 64 were classified as CWD+ (24 detected as positive at initial capture and 40 that transitioned to positive status during the study), while 56 were classified as CWD− as their final status. Locations were collected for an average of 441 days (range: 3 − 1,342 days) across all deer (Table S1). The cAE with 10%, 20%, 30% and 40% prevalence had 3:28, 6:26, 10:22 and 13:19 CWD+:CWD− animals for training. The logistic regression model showed limited discriminative ability between CWD− and CWD+ deer, with none of the predictors being statistically impactful (Table S2) and model testing resulting in an overall accuracy of 67%. The best performing threshold model was based on sinuosity and achieved an accuracy of 61%. In contrast, the AE with full holdout obtained 88% accuracy, followed by the AE with the 60/40 split with 84% accuracy (Fig. 1a). Accuracy was then highest at 82% for the cAE with 10% prevalence, followed by 67%, 65% and 52% accuracy for the cAE with 20%, 40%, and 30% prevalence, respectively (Fig. 1a). Precision was also highest in the AE with full holdout at 100%, followed by the AE with the 60/40 split at 89% (Fig. 1b). Precision was 63% for the cAE with 40% prevalence and precision declined with declining CWD prevalence (Fig. 1b). While TSS could not be calculated for the full holdout AE given no true negatives were in the testing dataset, the AE with the 60/40 split had the highest TSS with a value of 0.45, followed by the cAE models that all had overlapping ranges in TSS values across folds, with the average TSS ranging between − 0.09 and 0.21 (Fig. 1c). The cAE models correctly identified 0 of 3, 3 of 6, 4 of 10, and 5 of 13 of CWD+ deer for the 10%, 20%, 30% and 40% prevalence models, respectively. Fig. 1(a) Accuracy, (b) precision and (c) the true skill statistic (TSS) for the autoencoder model trained with a 60/40 split of CWD+/ CWD− animals, the autoencoder model trained with only CWD− animals (i.e., full holdout), and the conditional autoencoder models trained with 10%, 20%, 30% and 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA (a) Accuracy, (b) precision and (c) the true skill statistic (TSS) for the autoencoder model trained with a 60/40 split of CWD+/ CWD− animals, the autoencoder model trained with only CWD− animals (i.e., full holdout), and the conditional autoencoder models trained with 10%, 20%, 30% and 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA In the 60/40 AE and full holdout models, the models differentiated between CWD− and CWD+ deer using the rolling velocity standard deviation, absolute relative turn angle, rolling absolute relative turn angle skew, and sinuosity, with each contributing 10–15% of the total reconstruction errors (Fig. 2). In the cAE models, CWD− and CWD+ deer were differentiated by the lower bound of space use estimates in the 10% and 30% model, which made up 59.7% and 52.3% of the reconstruction error, respectively (Fig. 2). In the 20% model, differentiation was due to the lower bound of space use estimates (27.7% of the reconstruction error), rolling absolute relative turn angle mean (10.7%), and rolling velocity (10.0%, Fig. 2). In the 40% model, differentiation was due to the absolute relative turn angle (16.1%) and rolling velocity standard deviation (14.6%; Fig. 2). Collectively, there was a lack of consistency between cAE models in important variables (Fig. 2). Fig. 2Movement metric reconstruction error for the autoencoder model trained with a 60/40 split of CWD+/ CWD− animals (60/40), the autoencoder model trained with only CWD− animals (Full Holdout), and the conditional autoencoder models trained with 10%, 20%, 30% or 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA Movement metric reconstruction error for the autoencoder model trained with a 60/40 split of CWD+/ CWD− animals (60/40), the autoencoder model trained with only CWD− animals (Full Holdout), and the conditional autoencoder models trained with 10%, 20%, 30% or 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA For the AE models, false negatives showed inflated positive reconstruction errors for rolling absolute relative turn angle mean (8.5 and 9.6%) and rolling absolute relative turn angle skew (7.7 and 8.4%), while rolling velocity standard deviation was strongly reduced (-9.2 and – 8.9%; Fig. 3). In contrast, false positives did not show deviations from true positives above 5% (Fig. 3). There was also inconsistency across cAE models in the deviations that defined false negatives, with the 10%, 20% and 30% models having reductions in the lower bound of space use estimates (by -39.9 to − 18.2%) while the 10% model had greater errors for rolling absolute relative turn angle mean (by 13.9%), the 20% model had greater errors for rolling absolute relative turn angle standard deviation (by 10.1%), and the 40% model had greater errors for rolling velocity standard deviation (by 11.0%; Fig. 4). This was also true for false positives, with the 10% and 20% model having reductions in the lower bound of space use estimates (by -54.2 to -21.3%), the 40% model having reductions in the rolling velocity standard deviation (-12.0%), 10% model having greater errors for rolling velocity mean and standard deviation (by 13.8% to 21.6%) and the 20% model having greater errors for rolling absolute relative turn angle standard deviation (by 11.0%; Fig. 4). Fig. 3Deviations in movement metric reconstruction error from true positives and true negatives for (a) the autoencoder model trained with a 60/40 split of CWD+/CWD− animals and (b) the autoencoder model trained with only CWD− animals (i.e., full holdout) Deviations in movement metric reconstruction error from true positives and true negatives for (a) the autoencoder model trained with a 60/40 split of CWD+/CWD− animals and (b) the autoencoder model trained with only CWD− animals (i.e., full holdout) Fig. 4Deviations in movement metric reconstruction error from true positives and true negatives for the conditional autoencoder models trained with (a) 10%, (b) 20%, (c) 30% and (d) 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA Deviations in movement metric reconstruction error from true positives and true negatives for the conditional autoencoder models trained with (a) 10%, (b) 20%, (c) 30% and (d) 40% of animals being CWD+. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA Compared to CWD− deer, late-stage CWD+ deer exhibited reductions in space use across all space use metrics, with higher reconstruction errors in late-stage individuals compared to early-stage individuals (Fig. 5). Space use estimates were reduced by approximately 30–31% in late-stage individuals relative to CWD− deer. Velocity based metrics showed similarly strong declines, with early-stage CWD+ deer exhibiting substantial reductions in movement (Fig. 5), including decreases of approximately 26–29% in mean velocity and up to 44% decrease in velocity variability. In contrast, late-stage individuals showed slight increases in mean velocity (~ 5%) and velocity variability (~ 3%) relative to CWD− deer. Changes in turning behavior were also evident (Fig. 5). While early-stage individuals showed an increase in rolling absolute relative turn angle skew (~ 28% increase), late-stage individuals exhibited a decrease in turning attributes relative to CWD− deer (~ 20%). Fig. 5Movement metric % deviations between early-stage CWD+ and CWD− (Early-Neg), late-stage CWD+ and CWD− (Late-Neg), and early- and late- stage CWD+ (Early-Late) from the autoencoder model trained with only CWD− animals (i.e., full holdout), which was the best performing model. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA Movement metric % deviations between early-stage CWD+ and CWD− (Early-Neg), late-stage CWD+ and CWD− (Late-Neg), and early- and late- stage CWD+ (Early-Late) from the autoencoder model trained with only CWD− animals (i.e., full holdout), which was the best performing model. Models were for mule deer (Odocoileus hemionus) in Wyoming, USA Discussion The best performing models in our case study were the autoencoders that had either CWD− only (i.e., the full holdout model), or both CWD− and CWD+ animals (i.e., in a 40/60 split) in the training data, which achieved ≥ 84% accuracy and ≥ 89% precision, indicating they could consistently distinguish between CWD− and CWD+ deer from anomalous movement behavior. These results indicated that disease-related behavioral signals are not well captured by linear combinations of features or simple threshold rules, which are typical in models used in movement ecology to distinguish between animal states [45–47]. Instead, signals appear to be highly dimensional and nonlinearly expressed, limiting the effectiveness of conventional approaches. Additionally, the cAEs trained under cross validation consistently underperformed, with accuracies ranging from 54 to 77% and precision ≤ 53% across models. Autoencoders by their very nature are designed for efficient data compression and reconstruction, making them well suited for anomaly detection [38–43]. In this case study, when the AE was trained on a dataset containing a diverse and representative set of CWD− behavioral repertoires, it became highly efficient at identifying individuals that deviated from these learned behavioral signatures. In contrast, the cAE models were limited by their small training sample sizes, which restricted the cAE’s ability to learn what a CWD− baseline might look like across individuals with wide variation in behavioral patterns. Even as we increased CWD training prevalence, the cAEs continued to exhibit unstable feature importance, oscillating between dependence on variables related to space use or variables related to direction and sinuosity. This instability resulted in inconsistent signals and reduced generality. Together, these findings indicate that AEs can be reliable for anomaly detection but require sufficiently large sample sizes for training to do so. Therefore, if these types of models are to be used for CWD detection in wild populations of cervids, then managers will have to ensure they have collared a wide array of animals, and preferably prior to CWD detection on populations with low prevalence. Both AE models relied on rolling values related to reductions in velocity, absolute value of relative turn angles, and sinuosity, indicating that these models relied on a consistent set of combined movement cues related to speed and direction to distinguish between anomalous and normal movement patterns. Although the AEs performed well in determining CWD status through anomalous movement patterns, the use of these metrics resulted in some false positives, indicating that some CWD− deer exhibit bursts of movement or directional changes that mimicked disease-like behavior. By contrast, the AEs missed CWD+ deer whose anomalies were expressed through irregular turning patterns rather than through consistent changes in speed and sinuosity. These insights into how models identified misclassifications highlight how to adjust estimates to account for potential misclassifications when monitoring wild populations, thereby strengthening their value for management applications. Interestingly, early-stage CWD+ individuals exhibited stronger signals in behavioral shifts in movement speed and direction. Seemingly, deer infected with CWD begin displaying behavioral changes by reducing normal movement patterns even if they exhibit normal ranging behavior. In contrast, deer with late-stage CWD had massive reductions in space use, with little change to velocity and direction. The sequential pattern of declining speed and direction indicating a breakdown in locomotor functions, and then restrictive range use appears to highlight the progression of CWD, which has been documented in other studies [3, 18–21]. The ability to detect CWD-based anomalous behavior in various parts of progression of the disease indicates that animals are undergoing subtle movement-based changes throughout the infection period which presents an opportunity to use this information in disease detection approaches. For instance, AE models could be tuned to focus on separate sets of metrics, emphasizing space use metrics or direction and speed to recognize behavioral changes in animals no matter what stage of the disease they are in. Conclusions Our model demonstrates that AEs can effectively distinguish between CWD− and CWD+ mule deer based solely on GPS-derived movement features. In the age of biologging where many animals can be monitored remotely [48], the use of this tool has potential as an early warning system for wildlife managers. In recent decades, wildlife managers and researchers from a variety of institutions have deployed hundreds to thousands of GPS collars on cervids across North America and Europe for various research and management objectives. While disease surveillance may not be the primary goal of collar deployment, incorporating AEs to detect anomalous movement could not only provide additional coverage, but add to the justification to handle and equip animals with this technology. Our case study shows that we can leverage AEs to the task of disease monitoring for wildlife through examination of anomalous behaviors, but additional investigations into whether trained models can translate across populations or are generalizable to other cervids that are susceptible to CWD (i.e., white-tailed deer, Odocoileus virginianus) are needed. While AEs will not replace surveillance programs that directly test cervids in order to detect disease, they may provide an additional tool for early warning in areas where managers already collect relocation data on cervids. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1