预测警犬检测中的漏报情况。 机翻

Introduction

Detection dogs are widely used to identify target odors in tasks ranging from explosives detection to wildlife conservation and medical diagnostics1–5. Through training, dogs learn to make an indication response in the presence of a specific volatile chemical signature, allowing them to convert trace-level odor information into an observable indication for a human handler. In operational settings, these indications often guide critical decisions about the presence of hazardous substances, missing persons, or disease-related biomarkers. Nonetheless, like all biological systems, performance can fluctuate based on the vigilance, attentional and physical state of the detector6–9.

Operational errors in detection work occur in two principal forms: (i) false alerts in the absence of a target odor and (ii) missed alerts when a target is present but not indicated. While both types of error have operational implications, missed alerts are often of particular concern in high-stakes environments, where failure to detect explosives, hazardous substances, or disease-related odors may carry serious consequences. Evidence from canine detection work suggests that performance variability may be associated with changes in vigilance or attention, as well as with physiological demands imposed by ongoing physical work7,10–12.

One potentially unexplored explanation for variability in working dog performance lies in the physiological demands of detection work. Sustained locomotion, thermoregulatory load, and repeated odor discrimination tasks impose measurable physiological strain. Experimental studies have demonstrated that exercise intensity and duration can impair olfactory detection limits, particularly at low odor concentrations, even before overt behavioral signs of fatigue become apparent11. Alterations in autonomic regulation as reflected in heart rate dynamics and heart rate variability13–15, as well as increases in core body temperature16–20, may precede observable changes in search behavior. Together, these findings suggest that fluctuations in detection performance may arise from gradual shifts in physiological state. In humans, variability in task performance has been linked to changes in underlying physiological state, particularly in sustained attention paradigms21, and such fluctuations can manifest in subtle changes in movement patterns, as internal state modulates motor behavior including gait dynamics22.

In operational settings, the primary safeguard against performance decline is the handler’s ongoing assessment of the dog. Handlers monitor posture, gait, search intensity, response latency, and overall engagement, integrating these cues with knowledge of the individual dog’s training history and working style. Decisions to pause, rotate, or rest a dog are therefore grounded in expert judgment. This observational approach is indispensable and often effective. However, subtle physiological changes that precede overt behavioral alterations may not be readily visible, particularly in complex or high-pressure environments where attention is divided among navigation, communication, and safety considerations. Consequently, early stages of performance decline may remain difficult to detect.

Experimental research shows that workload and exercise influence autonomic regulation, thermoregulatory load, and movement dynamics—factors closely linked to olfactory performance11,12,23,24. These findings suggest that changes in quantifiable biological processes could be used as predictors of performance. Thus, identifying subtle behavioral changes associated with these physiological processes is critical for detecting early performance decline. The presence of these behaviors could be used by handlers to prevent odor misses before they happen.

To evaluate such signals, the temporal window relative to odor exposure must be precisely defined. In unconstrained search environments, the moment of first odor encounter is typically unknown, making it difficult to isolate behavior occurring before stimulus contact. Experimental investigation of pre-response signals therefore requires precise control over odor onset. To address this, we employed a controlled olfacto-treadmill system that enables computer-regulated odor delivery with precise temporal onset11. By combining an automated olfactometer with treadmill locomotion, dogs can be exercised under standardized conditions while receiving odor stimuli at known time points. The fixed lateral position of the dog further permits synchronized recording of video and physiological signals with minimal variability across trials and individuals. Heart rate and inter-beat intervals can be recorded continuously during locomotion, while core body temperature provides a complementary measure of physiological load. At the same time, the standardized camera view facilitates markerless pose estimation, allowing movement to be represented as the changing positions of anatomically defined body landmarks rather than through manual behavioral scoring. This controlled framework provides the temporal precision required to isolate and quantify pre-odor movement and physiological signals.

With this experimental precision in place, our goal was to test whether measurable variations in movement and physiology before odor exposure encode information about impending performance failure. Because posture is expressed through coordinated relationships among body regions and evolves continuously over time, we used supervised machine-learning models designed to capture both the anatomical structure of the dog’s movement and its temporal dynamics. We then asked whether these pre-odor patterns, alone and together with physiological measurements, could predict whether a subsequent target-present trial would result in a missed indication. Demonstrating such a relationship would indicate that information about subsequent performance is already present in the dog’s pre-odor behavioral and physiological state, rather than becoming apparent only after the trial outcome.

Results

Pre-odor movement predicts missed alerts

To assess whether impending detection failures are reflected in signals preceding odor contact, we analyzed data from a controlled olfacto-treadmill paradigm in which dogs performed repeated odor detection trials while moving at a fixed gait20 (Fig. 1). In this paradigm, the trained indication was a brief nose poke into the odor port, scored automatically when the muzzle interrupted an infrared beam for a predefined duration. Although this response differs from the sustained final responses commonly used in operational detection work, it provided an objectively measured indication outcome while allowing the dogs to maintain a standardized gait throughout the task. Dogs were tested both at walking (4 km/h) and trotting speed (8 km/h). Because performance decline was more pronounced under higher workload conditions, all analyses presented here were restricted to target-present trials performed at 8 km/h. We then analyzed the 4.5 s immediately preceding each target odor presentation. Behavioral outcomes (hit vs. miss) were used only as training labels.

[图：Markerless pose representation of a detection dog during treadmill-based odor detection. Full-body anatomical landmarks ]

Movement was quantified using markerless pose estimation25 to extract 29 anatomically defined landmarks per frame (Fig. 1). Each trial yielded a sequence of 135 frames (30 Hz), capturing joint-level posture and coordination over the pre-odor window. For each landmark, two-dimensional coordinates and a prediction confidence were recorded, resulting in a time series of 29 landmarks with three features each per frame. Coordinates were normalized relative to the withers to remove positional bias on the treadmill. This representation converts raw video into a time-resolved skeletal description of the dog’s posture and movement, enabling analysis of subtle shifts in head position, axial alignment, limb coordination, and tail dynamics.

Because anatomical landmarks form an articulated body structure, joints cannot be treated as independent variables. Instead, landmarks were represented as nodes in a graph reflecting skeletal connectivity, allowing the model to capture coordinated posture changes across the body. This graph representation was processed using a graph attention network (GAT), which enables the model to learn relationships between connected joints26. Movement patterns also evolve continuously over time rather than occurring in isolated frames. To capture these temporal dynamics, the sequence of pose graphs was processed by a temporal convolutional network (TCN), integrating movement information across the entire 4.5 s window27. Together, this spatiotemporal architecture (Fig. 2) allows the model to detect both brief micro-adjustments and slower drift in posture preceding odor exposure.

[图：Spatio-temporal models for predicting missed indications. Each video frame is converted into a skeletal representation d]

The pose-only model was trained using the recorded pre-odor sequences together with additional training examples derived from them by slightly shifting the analyzed time window and introducing small variations in landmark positions (augmentation). During training, missed indications and indications were represented equally. Model performance was then evaluated on an independent test set containing only original, unmodified trials and retaining the observed outcome distribution of the selected 8-km/h trials (216 trials: 144 missed indications and 72 indications). On this test set, pre-odor movement distinguished subsequent missed indications from indications with an AUC of 0.81 and an overall accuracy of 73.6%. Validation performance and training dynamics are provided in Supplementary Table S2 and Fig. S2. The model identified 77% of actual missed indications (recall = 0.77),and 82% of trials predicted as misses were indeed missed indications (precision = 0.82). Performance for indication trials was lower: 67% of actual indications were correctly identified (recall = 0.67), and 59% of trials predicted as indications were actual indications (precision = 0.59).

Because the model had no access to odor cues or trial outcomes at inference time, predictive performance must derive from movement structure expressed before stimulus contact. These findings demonstrate that subtle variations in posture and coordination during the pre-odor phase contain information about whether the subsequent target-present trial will meet the trained indication criterion.

Physiological signals enhance prediction of missed indications

Although movement thus contains predictive information, it reflects only the externally observable expression of behavior. Physiological measurements provide a complementary view of the internal state of the animal, including autonomic regulation and thermoregulatory load, both of which have been associated with performance variability in working dogs. We therefore examined whether incorporating physiological signals recorded during the same 4.5 s pre-odor window improves predictive performance.

To this end, we extended the pose-based model by adding three physiological inputs: heart rate (HR), inter-beat interval (IBI) statistics capturing short-term heart rate variability, and core body temperature. These signals were recorded concurrently with video and aligned to the identical pre-stimulus window. Unlike the pose data, which form a frame-wise time series across the 4.5 s window, physiological features were represented as summary measures over the interval and processed through a separate network branch. The resulting physiological representation was then fused with the movement-based embedding prior to classification (Fig. 2), allowing joint modeling of external movement dynamics and internal physiological state.

Incorporating physiological signals increased sensitivity to missed alerts while maintaining overall discrimination. On the independent test set, the multimodal model achieved an accuracy of 76.9% and an AUC of 0.80, comparable to the video-only model (AUC = 0.81). Validation performance and training dynamics are provided in Supplementary Table S2 and Fig. S3.

The primary effect of adding physiological information was an increase in the proportion of actual missed indications correctly identified by the model, from 77% to 85%. Precision for predicted missed indications remained essentially unchanged (0.82 versus 0.81), and the F1-score increased from 0.80 to 0.83. Overall discrimination was similar between the models (AUC = 0.81 for movement alone and 0.80 for the multimodal model). Performance for indication trials remained lower (precision = 0.67, recall = 0.61). Thus, the physiological inputs provided complementary predictive information primarily by improving recall for missed indications rather than by increasing overall AUC.

These findings show that short segments of autonomic and thermoregulatory data recorded before odor presentation provide complementary predictive information beyond movement alone. Incorporating physiological signals increases sensitivity to impending detection failures without materially affecting precision.

Behavioral and physiological features underlying prediction

Axial posture dominates movement-based prediction

While the preceding analyses demonstrate that pre-odor movement contains predictive information about missed alerts, they do not reveal which aspects of behavior change before a detection failure occurs. Identifying these behavioral signals is essential for interpreting the results in biological terms and for understanding what observable cues might indicate declining readiness in a working dog.

To identify which body regions carry this predictive information, we applied a gradient-based saliency analysis to the video-only model28. For each test trial, we quantified how small perturbations in individual joint coordinates affected the predicted probability of a missed alert and averaged these effects across frames and trials to obtain a mean saliency score for each landmark (Fig. 3A). Saliency values were rescaled for readability, with higher values indicating joints whose variation exerted a stronger influence on model output. A dynamic visualization of this saliency analysis is provided in Supplementary Video S2.

[图：Behavioral and physiological features underlying prediction. ( A ) Saliency ranking of pose landmarks derived from gradi]

Landmarks along the axial body line clearly dominated the saliency ranking. The top of the head (mean saliency: 0.212), the second cervical vertebra (axis; 0.132), and the withers (0.128) exerted the strongest influence on model output, indicating sensitivity to subtle changes in head and neck posture. Although the withers were used as the coordinate origin during preprocessing, the graph attention architecture propagates information from neighboring landmarks; saliency at this node therefore reflects the importance of relative axial posture encoded in the surrounding joints. The tip of the tail also ranked prominently (0.100), and the tip of the nose showed a moderate contribution (0.051). In contrast, most distal limb joints and thoracolumbar landmarks showed substantially lower influence (generally<0.034). One anatomically occluded landmark (right ischial tuberosity; 0.187) also exhibited elevated saliency. Because the dogs were filmed from the left side, this landmark was not directly visible and is therefore not interpreted as an observable behavioral feature. Its apparent importance may instead reflect unstable attribution under occlusion or indirect propagation of information from neighboring joints within the graph-based model. Accordingly, this landmark was interpreted with caution.

The concentration of predictive signal along the head–neck–trunk–tail axis is biologically plausible. Head posture directly reflects orientation toward the odor port and may encode subtle changes in engagement, attentional focus, or anticipatory sampling behavior during the pre-odor phase. Small shifts in head elevation or axial alignment may signal declining readiness before overt behavioral errors occur. Tail position and movement have been linked to arousal and emotional state in dogs29–31. That these regions emerged as dominant contributors without explicit anatomical weighting suggests that the model converged on behaviorally meaningful posture dynamics. Predictive information in pre-odor movement therefore appears to be concentrated in axial posture-related regions associated with engagement and arousal rather than distributed uniformly across the body.

IBI-derived features contribute most strongly among the physiological measures examined

To assess the contribution of individual physiological inputs, which form a small set of distinct modalities, we conducted a systematic ablation analysis32. In contrast to the pose representation, where landmarks are structurally interdependent within a spatiotemporal graph, physiological signals can be meaningfully removed at the modality level, enabling direct evaluation of their effects. Each physiological modality was removed in turn while all other inputs were retained, and predictive performance was reassessed on the test set using AUC, classification accuracy, and log-loss for evaluation (Fig. 3B).

Removing IBI-derived features produced the largest degradation in predictive performance. The area under the ROC curve (AUC) decreased by 0.16 (95% CI: 0.08–0.23), accuracy dropped by 6 percentage points, and log-loss increased by 0.19 relative to the full multimodal model. Eliminating core temperature resulted in a more moderate performance decrease (

[内容请查看原文]

AUC = 0.06, 95% CI 0.03–0.10; accuracy

[内容请查看原文]

%; log-loss

[内容请查看原文]

). In contrast, removing heart rate alone had minimal impact on discriminative performance: AUC and accuracy remained essentially unchanged, although log-loss increased (

[内容请查看原文]

log-loss = 0.49), indicating reduced probability calibration without a shift in the decision boundary.

Among the physiological inputs examined, short-window IBI-derived features provided the largest contribution to prediction. These features summarize beat-to-beat cardiac timing within the 4.5-s pre-odor interval and should therefore not be interpreted as conventional longer-duration HRV estimates. Beat-to-beat cardiac timing is influenced by autonomic regulation, but it is not a process-specific measure of autonomic state. Variation in these features may reflect interacting influences related to arousal, cardiovascular and thermoregulatory regulation, sensory-reflex processes, or broader physiological state, which the present analysis cannot distinguish. Core temperature provided additional predictive information, whereas heart rate alone contributed little to discrimination within the present feature set. Together, these findings show that short-window cardiac timing dynamics provide the strongest physiological predictive signal among the measures examined and that this signal is present before odor contact

Converging markers of declining readiness

Taken together, the behavioral and physiological analyses successfully predict impending misses. Predictive information in behavior is concentrated in head posture and tail dynamics—external markers of engagement and arousal—while predictive information in physiology is dominated by short-term heart rhythm variability, an indicator of autonomic regulation. The convergence of these signals suggests that declining detection performance is not triggered by a single behavioral event but reflects a broader change in the dog’s physiological and attentional state. Importantly, these markers emerge within the 4.5 s window preceding odor presentation, before the dog has encountered the stimulus. By linking predictive performance to specific behavioral and physiological features, the framework moves beyond opaque classification toward a biologically interpretable description of early performance decline.

Aggregated model outputs illustrate evolving miss risk within a session

Single-trial predictions fluctuated and therefore did not directly capture gradual changes in performance within a session. We therefore examined whether trial-level model outputs could be aggregated to approximate an evolving performance state across consecutive trials. Model confidence scores were treated as probabilistic evidence and sequentially integrated across target-present trials using a Beta-based accumulation framework. This procedure yields a continuously updated estimate of the underlying miss probability, producing a smooth trajectory of inferred miss risk over time (Fig. 4A).

[图：Aggregated model outputs track evolving miss risk within a session. ( A ) Conceptual illustration. Trial-level predictio]

As an illustrative example, we analyzed a complete run from one dog comprising 50 target-present trials (Fig. 4B). For each trial, the model generated a confidence score based solely on the 4.5 s pre-odor window. These scores were aggregated sequentially to estimate the probability that the underlying miss rate exceeded predefined thresholds. This trajectory was compared to the observed miss frequency calculated over sliding windows of 10 trials. As the empirical miss frequency increased during the session, the aggregated model output increased in parallel. Because model predictions were generated without access to trial outcomes, this correspondence reflects information contained in pre-odor movement and physiological signals.

In this example, a neutral prior was used, corresponding to no strong initial assumption about the dog’s miss rate. Alternative prior specifications produced qualitatively similar trajectories (Fig. S4).

These observations demonstrate that aggregating trial-level predictions can capture gradual changes in performance within a session. However, this analysis is based on a single illustrative example and does not constitute a systematic evaluation across dogs or conditions.

Discussion

A central question motivating this study was whether fluctuations in detection performance reflect an underlying biological state that becomes detectable before an error occurs. Using a controlled olfacto-treadmill system that allowed precise isolation of the seconds preceding odor exposure, we found that movement patterns alone contained sufficient information to predict subsequent missed alerts. Predictive sensitivity increased further when physiological signals were incorporated. Because all input data were recorded strictly before odor delivery and without access to behavioral outcomes, the predictive signal must arise from variations in the dog’s behavioral and physiological state rather than from reactions to odor processing. These findings show that missed indications are preceded by a measurable configuration of behavior and physiology, supporting the view that detection performance reflects a dynamic performance-related state rather than a sequence of isolated stochastic failures.

The predictive signal was not diffuse across the body but concentrated in anatomically and behaviorally meaningful regions, particularly cranial landmarks and the tip of the tail. Subtle changes in head posture may reflect altered attentional orientation or sampling readiness during the pre-odor phase, whereas tail dynamics are widely associated with arousal and motivational state in dogs29–31. The prominence of these regions suggests that missed alerts are preceded by subtle deviations in posture rather than by gross locomotor disruption. Physiological measurements provided complementary information, with short-window IBI-derived features contributing more strongly than raw heart rate or core temperature among the modalities examined. Interpretation of these features is constrained by the 4.5-s analysis window: although they captured predictive variation in beat-to-beat cardiac timing, they should not be considered equivalent to conventional HRV measures derived from substantially longer recordings. Future studies could extend the present approach by analyzing longer continuous cardiac recordings between odor presentations, allowing established HRV measures to be related to both short-term predictions and longer-term changes in performance. This would be particularly informative because the stronger contribution of beat-to-beat timing compared with raw heart rate is consistent with the possibility that regulatory dynamics, rather than workload reflected simply by an elevated heart rate, contribute to the predicted state. Longer-window HRV measurements could help test whether this signal is associated with broader changes in autonomic regulation related to arousal, engagement, or sustained task demands13,14.

Importantly, these alterations emerged while the dogs continued trotting under constant workload, without visible disruption of gait or stability. This indicates that missed alerts are not preceded by overt behavioral breakdown but by more subtle perturbations in the biological systems supporting sustained detection work. Detection tasks—particularly at low odor concentrations—place continuous demands on attentional and regulatory processes. Under such conditions, even modest changes in physiological or cognitive state may influence the probability of a successful indication.

If detection performance fluctuates as a biological state, the question becomes how such changes should be interpreted during ongoing work. In practice, handlers are less concerned with isolated events than with the dog’s current state. The relevant question is therefore: What is the likelihood that my dog is operating in a condition where missed alerts exceed an acceptable threshold? This is the level at which deployment decisions—continue searching or allow recovery—are typically made.

Interpreting model outputs at this level therefore requires moving beyond single-trial predictions toward an estimate of the dog’s evolving performance state. In the present study, we aggregated consecutive trial-level predictions to estimate whether the dog’s underlying miss rate was increasing beyond an acceptable threshold. By integrating model outputs across trials, this approach translates fluctuating single-trial predictions into a probabilistic assessment of performance state. Such an estimate can incorporate prior information, including the dog’s historical performance or the expected difficulty of the task, while the miss-rate threshold of interest can be adjusted to reflect the consequences of a missed indication. In this way, the Bayesian aggregation provides a structured, context-sensitive framework for translating trial-level predictions into an evolving estimate of miss risk (Fig. 5). In the present study, we primarily present the statistical framework for this aggregation and illustrate its application using a single session. This limitation concerns the aggregation layer; the underlying trial-level predictive models were evaluated on the independent held-out test set. This illustration does not constitute systematic validation of the aggregation layer or establish operational decision thresholds; such validation will require evaluation across dogs, sessions, task conditions, and field environments.

[图：Conceptual framework for translating pre-response behavioral and physiological signals into estimates of performance sta]

The controlled conditions that made precise isolation of the pre-odor interval possible also differ substantially from operational searches. In particular, odor was delivered from a fixed, visible port at a known location. The dogs therefore did not need to locate or source an unknown target, as would be required during free-search detection work. Operational searches additionally involve heterogeneous odor plumes, variable terrain, handler interaction, environmental distractions, and uncertain target locations. The present paradigm11 deliberately removes these sources of variability to provide precise stimulus timing; whether the identified pre-odor signatures remain predictive when these additional search demands are introduced requires direct field validation.

The response requirement itself represents a second difference from many operational detection tasks. In the olfacto-treadmill paradigm, dogs were trained to perform a brief nose-poke indication, whereas operational dogs may instead be required to sustain a sit, down, stand-and-stare, or other final response. Because all model inputs ended at odor onset, the response itself cannot generate the anticipatory signal reported here; however, the response requirement defines which trials are subsequently classified as indications or missed indications. Indication duration can alter response bias and response type can influence aspects of sample-checking and hesitation behavior33,34. The brief nose-poke criterion should therefore be regarded as the trained indication specific to the present paradigm, and validation across sustained operational response types will be required to establish the generality of the pre-odor predictors.

The present sensor configuration should be viewed as an experimental measurement system rather than as an operational monitoring package. This limitation applies both to the physiological measurements and to the video-based pose estimation25: continuous, unobstructed lateral video is rarely feasible during operational searches. However, the present findings suggest that the predictive information resides in movement structure rather than in the video modality itself. Wearable inertial sensors and accelerometers are already widely used in canine research and may provide a more field-compatible means of capturing analogous movement signatures35–37. Whether the physiological information identified here can likewise be recovered reliably using sensors suitable for unrestricted detection work remains to be tested. Translating the present findings into an operational monitoring system will therefore require validation of both movement and physiological measurements under active field conditions.

A related consideration concerns the temporal scope of the analyzed signals. The present study focused on short pre-odor intervals defined by the experimental structure of the task, in which odor presentations occurred at fixed time points. This design enabled precise isolation of anticipatory signals while minimizing contamination from preceding trial outcomes, which themselves can influence the dog’s immediate behavioral and physiological state. As a result, the model operates on temporally localized snapshots of the dog’s condition immediately prior to each odor encounter.

It is notable that predictive information was detectable within this brief window, indicating that measurable markers of declining readiness can emerge on a sub-trial timescale. At the same time, performance fluctuations are likely to evolve over longer periods spanning multiple trials. Future work may therefore examine how signals at different temporal scales—from short anticipatory intervals to longer-term trends—can be combined. Importantly, such approaches would complement rather than replace the present framework, in which aggregation of short-window predictions provides a mechanism for tracking gradual changes in performance state across a session.

Future mechanistic studies could complement the present behavioral and physiological measures with neural measurements such as non-invasive EEG or fNIRS in suitably adapted controlled paradigms. Combined with experimental manipulation of attentional demand, motivation, or workload, such measurements could help distinguish whether the predictive pre-response state is associated specifically with attentional processing or reflects broader performance-related changes. The olfacto-treadmill therefore provides a useful controlled platform for identifying candidate signals and mechanisms before testing them in progressively more complex environments. By allowing precise experimental control, it enables identification of candidate behavioral and physiological markers of declining readiness before testing them in progressively more complex environments. Future work should evaluate whether similar anticipatory patterns can be detected across larger populations, diverse task types, and operationally realistic conditions.

In summary, we show that missed alerts in this controlled detection task are preceded by measurable variation in posture and physiology. This indicates that information about subsequent performance is detectable immediately before a missed alert occurs. Detection performance therefore appears to reflect an evolving biological state rather than a sequence of isolated successes and failures. While further validation will be required before operational deployment, these findings suggest that monitoring such signals may enable earlier and more informed decisions about whether a dog should continue working or be allowed to recover, with implications for both operational reliability and animal welfare.

Methods

Ethics statement

The study was conducted at the Texas Tech University Canine Olfaction Research and Education (CORE) Lab and approved by the Texas Tech University Institutional Animal Care and Use Committee (IACUC) and the U.S. Army Animal Care and Use Review Office (ACURO). All methods were performed in accordance with the relevant guidelines and regulations.

Data collection and preprocessing

Subjects

Four dogs participated in this study (Table 1). All dogs were between 2 and 4 years of age, healthy, and actively trained for detection work. One dog (Charles) was the pet dog of a laboratory staff member and was brought to the facility daily for training and testing. The remaining three dogs (Adele, Boomer, and Iitooma) were procured for the project from Chiron K9 (Somerset, TX, USA) and were housed at the Texas Tech University Canine Olfaction Research and Education (CORE) Lab facility. One dog (Charles) had also participated in our prior treadmill study11. Dogs were selected based on their familiarity with treadmill walking and consistent participation across sessions. This information has previously been reported in a companion manuscript20.

Experimental setup and data acquisition

Olfacto-treadmill setup and task design All recordings were obtained during a treadmill-based odor discrimination task using a custom-built olfacto-treadmill11,20. Dogs maintained a consistent gait—either walking (4 km/h) or trotting (8 km/h)—throughout the session, while odors were delivered via an automated olfactometer. All analyses in the present study were restricted to target-present trials performed at 8 km/h.

Trial structure Each trial began with computer-controlled activation of an odor valve (target or diluent). Dogs were trained to perform a nose-poke indication by inserting the muzzle into the odor port and interrupting an infrared beam. On target-present trials, an indication was scored only when the beam was interrupted for the predefined minimum duration: 0.4 s for Boomer and Charles and 0.8 s for Adele and Iitooma. These dog-specific thresholds were established during training to accommodate differences in muzzle shape and reliable interruption of the infrared beam, rather than differences in detection performance, and remained fixed during experimental testing. A missed indication was scored when no criterion-duration response was completed within the 10-s response period; brief subcriterion beam breaks did not by themselves constitute an indication. Correct criterion responses were reinforced with food. Trials were initiated every 20 s, yielding fixed onset intervals and uniform session durations. Target presence was randomized at 50% frequency. The experimenter was blind to trial type; valve selection and feedback tones were fully automated. The target odor was 1-bromooctane (1-BO) diluted in food-grade mineral oil. Target concentration was determined for each dog by threshold testing conducted before data collection and then held constant during experimental testing:

[内容请查看原文]

v/v for Adele and Boomer and

[内容请查看原文]

v/v for Charles and Iitooma. Filtered air passing through the odor vial at 2 L/min was mixed with a continuous 8 L/min carrier-air stream, corresponding to a 20% odor-line air dilution. Airflow settings remained constant across all analyzed trials.

Session formats. Two session types contributed data: (i) Fixed-pace sessions included 100 trials at a constant gait (walk or trot), lasting 33 min 20 s. (ii) Pace-transition sessions included 140 trials (approximately 46 min) with a mid-session gait switch (walk

[内容请查看原文]

trot or trot

[内容请查看原文]

walk) after trial 70. Each dog completed four fixed-pace sessions (two walk, two trot) and eight pace-transition sessions (four of each transition type) on separate days, with at least 18 h between sessions. Exercise was voluntary; if a dog stepped off the treadmill, the session resumed once the dog re-engaged.

Video recording Video was recorded from a fixed lateral camera positioned at mid-body height under consistent indoor lighting and background. Recordings were time-synchronized with the olfactometer and physiological event logs. For each included trial, the 4.5 s immediately preceding odor onset (defined by the olfactometer valve activation timestamp) were extracted for modeling. Details of pose estimation and feature extraction are provided in the following section.

Physiological measurements Core body temperature was measured using an ingestible AniPillTM capsule (BodyCap, France), sampled every minute. Heart rate (HR) and inter-beat intervals (IBI) were continuously recorded using a Polar® H10 chest strap and logged at 1 Hz. All physiological data were aligned to trial timing. Temperature values were temporally aligned to trial onset using the closest preceding measurement.

Subset used for modeling To isolate predictors of missed indications under moderate workload, analyses were restricted to target-present trials at 8 km/h. Trials from both fixed-pace and pace-transition sessions were pooled for analysis, as all included trials were performed at the same gait (8 km/h) and extracted using an identical pre-odor window. For each included trial, video and physiological data were aligned to the 4.5 s pre-odor window, and behavioral labels (hit vs. miss) were assigned based on the infrared beam-break criteria described above.

Physiological data processing

Gastrointestinal temperature data were sampled at 1 min intervals and linearly interpolated to 1 s resolution. Heart rate (HR) was recorded continuously at 1 Hz. All physiological signals were time-aligned to olfactometer valve activation timestamps defining odor onset. For each included trial, physiological data were extracted from the 4.5-s interval immediately preceding odor onset. Only pre-stimulus data were used for feature extraction. Trials containing signal loss, sensor dropout, or incomplete physiological recordings within the pre-odor window were excluded from analysis.

Trial labeling and outcome definition

Each retained target-present trial was labeled according to whether the dog completed the trained criterion indication during the response period. Trials meeting the criterion were labeled as indications, whereas trials in which no criterion-duration response was completed were labeled as missed indications. These labels describe observable task performance and do not establish whether the target odor was perceived in the absence of a criterion response. Behavioral labels were used exclusively for model training and evaluation; no post-odor outcome information was available during extraction of the pre-odor movement or physiological features.

Data preprocessing and inclusion criteria

Only target-present trials with complete and valid pose and physiological recordings were included in the analysis. For each retained trial, a 5 s interval immediately preceding odor onset (defined by the olfactometer valve activation timestamp) was initially extracted.

From this 5 s segment, a 4.5 s window was used for model input. For non-augmented samples, the final 4.5 s immediately preceding odor onset were selected. For augmented samples (see Supplementary Section S2), a randomly shifted 4.5 s subwindow within the original 5 s segment was extracted to reduce alignment bias across trials. All features were aligned to the selected 4.5 s window prior to preprocessing.

Pose data The initial 5-s segment contained 150 frames sampled at 30 Hz. The final model input consisted of 135 consecutive frames (4.5 s). Pose annotations included 29 anatomical landmarks per frame (see Fig. 1), each represented by two pixel coordinates (x, y) and a likelihood value. To remove global positional bias, coordinates were translated such that the withers (joint 4) were located at the origin in each frame. This normalization reduced sensitivity to absolute body position on the treadmill and within the video frame, emphasizing relative joint movement.

Heart rate Heart rate was sampled at 1 Hz. Values corresponding to the selected 4.5 s window were extracted from the original 5 s segment. Missing values due to brief signal loss were linearly interpolated.

Inter-beat interval (IBI) Inter-beat intervals within the selected 4.5 s window were used to compute seven summary statistics: (1) mean RR interval, (2) standard deviation, (3) minimum, (4) maximum, (5) root-mean-square of successive differences (RMSSD), (6) coefficient of variation, and (7) mean absolute successive difference. These statistics were treated as short-window IBI-derived descriptors for prediction. Trials containing fewer than two RR intervals within the window were excluded.

Core temperature For each trial, the most recent temperature measurement at or preceding odor onset was selected.

All features (pose coordinates, heart rate, IBI statistics, and temperature) were z-scored per dog to reduce inter-individual differences in baseline physiology and body size.

Pose estimation and segmentation-assisted refinement

Movement was quantified using markerless pose estimation to obtain a time-resolved skeletal representation of each dog during treadmill locomotion. We trained a DeepLabCut-based model25 to track 29 anatomical landmarks spanning the entire body, including cranial reference points, axial markers along the spine, forelimb and hindlimb joints, and tail landmarks (Fig. 1). The landmark set was designed to provide comprehensive anatomical coverage without imposing region-specific weighting or prior assumptions regarding which body segments might be predictive.

A total of 120 frames were manually annotated across dogs and sessions to capture a broad range of postures and movement phases. A single model was trained and subsequently applied across all individuals. For each frame (30 Hz), the network produced two-dimensional landmark coordinates and associated likelihood values.

To improve robustness under varying background conditions and partial occlusion, pose estimation was performed on segmentation-refined frames (details in Supplementary Section S1; validation in Supplementary Fig. S1 and Table S1). An initial pose prediction was used to define a region of interest centered on the dog’s body. Within this region, a segmentation model was applied to isolate the dog from the background38, and pose estimation was repeated on the masked image. The region of interest was updated dynamically across frames based on the preceding pose prediction, enabling stable tracking throughout each sequence.

All pose sequences used in downstream analyses were derived from these segmentation-refined predictions. A visual example of the segmentation-assisted refinement procedure, including landmark overlay and background removal, is provided in Supplementary Video S1.

Model architecture and training

Spatiotemporal graph network (ST-GAT-TCN) Pose-only model. To model structured movement dynamics, we implemented a spatiotemporal graph neural network (ST-GAT-TCN)39 (Fig. 2). Each input sample consisted of a sequence of 135 consecutive frames (4.5 s at 30 Hz), with 29 anatomical landmarks per frame. For each landmark, two-dimensional coordinates (x, y) and a confidence value were used, yielding an input tensor of shape (135, 29, 3).

Spatial dependencies between joints were modeled using a fixed undirected graph with 29 nodes, corresponding to the anatomical landmarks. Edges reflected musculoskeletal connectivity, including spinal chains and limb–torso attachments. This graph structure was applied consistently across all time steps. Frame-wise spatial processing was performed using stacked graph attention layers, enabling adaptive weighting of neighboring joints while preserving anatomical structure.

Temporal dynamics were modeled using a multi-layer temporal convolutional network (TCN) operating along the time axis. Dilated one-dimensional convolutions were employed to capture movement patterns across multiple temporal scales within the 4.5 s pre-odor window. The resulting temporal representation was globally pooled and passed through fully connected layers, with a final sigmoid output producing the predicted probability of a missed indication.

Multimodal model with physiological input. To incorporate internal physiological state, the pose-based ST-GAT-TCN architecture was extended with a parallel physiological branch (Fig. 2). For each trial, heart rate, inter-beat interval (IBI) summary statistics, and core body temperature extracted from the 4.5-s pre-odor window were used as additional inputs.

Physiological features were processed through dedicated feedforward layers to generate a compact physiological embedding. This embedding was combined with the spatiotemporal pose representation via late fusion prior to the final classification layer. The fused representation was passed to the same sigmoid output head used in the pose-only model.

This modular design preserved the spatiotemporal backbone of the pose-based network while enabling joint modeling of behavioral movement dynamics and autonomic physiological state. Additional architectural and training details are provided in Supplementary Section S2.

Training and evaluation

Model development was restricted to target-present trials with complete pose and physiological data available within the 4.5 s pre-odor window. Trials were first partitioned into a held-out test set and a development set. The test set was not used during model training, augmentation, or model selection and contained only original, unmodified samples. The independent test set was not class-balanced and retained the observed outcome distribution of the selected trials.

Within the development set, data augmentation was applied to increase robustness and reduce sensitivity to fixed temporal alignment within the pre-odor window. For each trial, additional samples were generated by extracting randomly shifted 4.5-s subwindows from the original 5-s clip and applying small spatial perturbations to joint coordinates. Augmentation was performed separately within each class, and class balancing was achieved by matching the number of samples across indication and miss trials. Class balancing was therefore applied only to the development data.

The resulting development data were then randomly partitioned at the sample level into training and validation subsets. Model selection was based on validation AUC. Each architecture was trained multiple times with different random initializations to ensure stability of performance estimates. Run-wise validation metrics are summarized in Supplementary Table S2.

Final model performance was reported exclusively on the independent held-out test set using accuracy, area under the receiver operating characteristic curve (AUC), precision, and recall.

Interpretability analyses

Saliency analysis of pose features. To identify which anatomical landmarks most strongly influenced movement-based predictions, we analyzed the trained pose-only model using a gradient-based saliency approach28. For each trial in the held-out test set, the gradient of the model output with respect to each joint’s input features (two-dimensional coordinates and detection confidence) was computed in evaluation mode without updating model weights. Absolute gradient magnitudes were averaged across frames within the 4.5 s pre-odor window and subsequently averaged across trials to obtain a global importance score per landmark.

Physiological sensor ablation. The contribution of physiological signals was assessed using a sensor ablation procedure32. For each physiological modality (heart rate, inter-beat interval features, and core body temperature), inputs from the corresponding branch were replaced with zeros while all other inputs were left unchanged. Using the same trained multimodal model, predictions were recomputed on the held-out test set and compared against the full-input condition using AUC, accuracy, and log-loss.k

Bayesian aggregation of trial-level predictions

To approximate an evolving performance state across consecutive trials, we modeled the dog’s underlying miss probability within a session as an unknown latent parameter

[内容请查看原文]

. Within a given session,

[内容请查看原文]

was assumed constant but updated sequentially as probabilistic evidence accumulated from trial-level model outputs. A Beta prior distribution,

[内容请查看原文]

, was specified to represent baseline uncertainty regarding miss probability at session onset. Unless otherwise stated, a weakly informative prior was used (

[内容请查看原文]

), corresponding to a uniform distribution over [0, 1]. Further mathematical details and prior-sensitivity analyses are provided in Supplementary Section S3.

For each target-present trial i, the trained model generated a predicted probability

[内容请查看原文]

of a missed alert based exclusively on the 4.5 s pre-odor window. These predicted probabilities were treated as fractional Bernoulli observations and used as soft evidence to update the Beta posterior sequentially according to:

[内容请查看原文]

This update yields a closed-form posterior distribution

[内容请查看原文]

after i trials. At each step, the posterior mean miss probability was computed as:

[内容请查看原文]

In addition, exceedance probabilities were calculated to quantify the likelihood that the underlying miss rate exceeded a predefined operational threshold

[内容请查看原文]

. Specifically,

[内容请查看原文]

was computed analytically from the cumulative Beta distribution. To facilitate comparison with observed behavioral performance, empirical miss frequency was calculated using a trailing sliding window of 10 consecutive target-present trials. Aggregated posterior trajectories were then compared qualitatively to the sliding-window miss rate within the same session.

Supplementary Information

[内容请查看原文]

[内容请查看原文]

[内容请查看原文]

Acknowledgements

Funding was provided by the U.S. Army via the Chemical Biological Advanced Materials and Manufacturing Science (CBAMMS) Program (PE 0601102A Project VR9) at the Combat Capabilities Development Command Chemical Biological Center (DEVCOM CBC).

Author contributions

Conceptualization: Jörg Schultz, Edgar Aviles-Rosa, Nathaniel J. Hall, Michele N. Maughan Data curation: Jörg Schultz Formal analysis: Jörg Schultz funding acquisition: Michele N. Maughan Investigation: Liza Rothkoff, Edgar Aviles-Rosa, Nathaniel J. Hall Methodology: Jörg Schultz, Liza Rothkoff, Edgar Aviles-Rosa, Nathaniel J. Hall, Michele N. Maughan Project administration: Michele N. Maughan Software: Jörg Schultz Visualization: Jörg Schultz Writing—original draft: Jörg Schultz Writing—review and editing: Jörg Schultz, Liza Rothkoff, Edgar Aviles-Rosa, Nathaniel J. Hall, Michele N. Maughan.

Data availability

All code, videos, trained models, and example data required to reproduce the analyses are available at Zenodo: https://doi.org/10.5281/zenodo.17197669 (version v2).

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

[内容请查看原文]

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-71367-8.

References

[内容请查看原文]

Associated Data

Supplementary Materials

[内容请查看原文]

[内容请查看原文]

[内容请查看原文]

Data Availability Statement

All code, videos, trained models, and example data required to reproduce the analyses are available at Zenodo: https://doi.org/10.5281/zenodo.17197669 (version v2).