We recruited 55 healthy adults (27 males and 28 females aged 18–59 years). Exclusion criteria were pre-existing respiratory tract infections or other underlying health conditions. Additionally, we only recruited participants who did not intend to receive a COVID-19 or seasonal influenza vaccine during the study period, and participants had to be clear of any other inflammation within 7 days prior to the start of the study. Data were collected from December 2021 to February 2022. Participants were recruited through word-of-mouth and posted advertisements in Montreal, Canada. Further details about the inclusion, exclusion, and withdrawal criteria are provided in the study protocol (Hadid et al., 2023).

The study period lasted 12 days and was divided into two distinct phases: the pre-inoculation period (considered as day −7 to day 0) and the post-inoculation period (day 0 to day 4). Starting from day −7, participants continuously wore a Astroskin (10 participants)/Hexoskin (45 participants) garment (Carré Technologies, Inc. Montreal, Canada) to monitor their physiological markers (Figure 1A). Blood samples were also collected (Figure 1B), 3-min CRSST followed by 2 min of recovery was conducted (Figure 1C). The Astroskin/Hexoskin garment was used to collect the following physiological markers relevant to respiratory health:

Heart Rate (HR): The number of heartbeats per minute (bpm)

The data format of the measured physiological markers is a 1-D array counting 300 rows (1 data point every second) per physiomarker per visit. Each data point is an integer number for HR, BR, and MV, and a floating point value for RRI. Although the Astroskin shirt also measures blood oxygen concentration, we limited analyses to signals that were common to both garments (the exact same raw data is provided by both biosensors). Prior validation studies have demonstrated that both devices offer highly comparable and reliable measurements for the physiological markers considered in this study (Harbour et al., 2021; Montes et al., 2018; Smith et al., 2019). The choice of these devices was driven by their non-invasive design, ability to capture continuous multi-channel physiological data, and validated accuracy in measuring cardiac and respiratory signals during physical activity and recovery. On day 0, blood samples, and a 3-min CRSST at 30 steps/min for all participants were performed to provide additional baseline data and assess participants’ physiological responses during exertion. Participants were then inoculated with a live attenuated influenza vaccine (LAIV): FluMist® (Carter and Curran, 2011). The post-inoculation period (5 days) involved daily blood sample collection to monitor the participants’ response to the inoculated influenza vaccine. Participants performed the 3-min CRSST three additional times during the post-inoculation period (days 1, 2 and 3).

Blood samples facilitated the measurement of biomarkers and immune-related factors, aiding in identifying baseline levels and subsequent changes induced by the influenza vaccine. The inflammation level was obtained by calculating the immune response score (IRS) which is the objective of our prediction model. This score is composed of the mean value of 8 measured biomarkers (IFN γ , IL-6, IL-8, IL-10, IL-15, IP-10, MCP-1, and TNF α ) during each visit (Equation 1). A detailed definition of each biomarker is presented in Table 1. These biomarkers were chosen due to their ability to detect inflammation in the host immune system (McClain et al., 2015; Davey et al., 2013; Lee et al., 2011; Chi et al., 2013; Peiris et al., 2004; Monteerarat et al., 2010). We then used 4 IRS scores obtained from the baseline period (2 scores per visit), to calculate the IRS’ coefficient of variance (CoV) for each participant (Equation 2), allowing us to assess the individual variation of the IRS values. We used these CoV values to apply a 0-CoV normalization (Equation 3) to the original IRS. By normalizing the IRS values, the individual baseline variations among all participants are eliminated, thus, providing generalizable scores of inflammation levels. Additional analyses using the raw IRS scores (without normalization) were performed to assess the impact of this step (Table 2).

The stair test consists of stepping for 3 min (with a consistent pace of 30 steps/minute) up and down a step having a height of 20 cm (Hadid et al., 2023). We chose this controlled clinical test as it’s been previously used to uncover information about a participant’s physiological response to exercise-induced stress (Li et al., 2021; Elsaid, 2011; Kieu et al., 2020; Lee et al., 2019). Following the test, participants recovered in a seated position for 2 min. Physiological markers were monitored during the entire session.

Although oxygen consumption was not directly measured during the 3-min CRSST, exercise intensity can be estimated using the American College of Sports Medicine (ACSM) stepping equations. Based on the fixed stepping rate (30 steps/min), step height (20 cm), and individual body mass, the ACSM equations allow estimation of oxygen consumption and corresponding metabolic equivalents (METs) at the end of the CRSST. These estimates indicate that the CRSST elicits a moderate-intensity workload, approximately 5–7 METs, corresponding to 50%–70% of maximal oxygen uptake ( VO 2 max ) in healthy adults (Lewthwaite et al., 2020).

First, the 5-min period of interest (stair test and recovery) was segmented based on annotated timestamps, resulting in 238 visits having 4 signal data files per visit and 300 data points (integer numbers) per signal (Figure 3A). We then plotted all the sensor data for each visit to assess the signal quality and identify missing or noisy segments through a visual inspection (Figure 3B). Examplars of noisy and typical heart rate signals are shown in Figure 2. During this step, data were removed due to noisy signals from BR (24 visits), missing or noisy HR signal (49 visits), noisy MV signals (3 visits), and missing or noisy RRI signal (68 visits). Next, we performed data cleaning. This involved identifying inconsistencies and errors in the dataset that were flagged by Hexoskin’s systems and the data cleaning protocols (detailed in Table 3 (Lewthwaite and Jensen, 2021; Blackie et al., 1991; Foll et al., 2021) and Figure 3C). The data points having such inconsistencies and errors are discarded from the record of the concerned visit (n = 10). Finally, we performed artifact removal for the RRI data following 3 rules (Acar, Karlsson and Malik rules) which resulted in excluding 4 visits. (Foll et al., 2021) (Figure 3D). It is worth noting that some visits met multiple exclusion criteria. The final number of remaining visits is 170. We extracted the HRV features during the first minute of recovery using the Flirt Python package (Foll et al., 2021). Flirt offers tools and algorithms specifically designed for extracting HRV-related features from the RRI signal, including statistical (e.g., min, max, mean), time-domain (e.g., skewness, kurtosis, number of peaks), and frequency-domain features (e.g., low frequency to high frequency ratio) (Figure 3E). Furthermore, we calculated additional features to further enhance the analysis (Figure 3F). These features specifically targeted the BR, HR as well as the MV signals, and included the increase or decrease from the baseline during specific time intervals of the stair test, such as minutes 0 to 3 and 3 to 4 (Equation 4). An AUC feature was also calculated (using the Trapz function provided in the Numpy Python package (Developers, 2025)) for the 3 min of activity and the subsequent minute of recovery. A full list of features is provided in Table 4. Finally, we concatenated the cleaned data, extracted and calculated features, and consolidated all relevant information for a particular visit into a cohesive record. Afterwards, we merged all visits together resulting in a final dataset having 170 rows (visits) and 40 columns (39 features and the corresponding IRS score for each visit) (Figure 3G).

Selecting an appropriate model for our task is a critical step. Our objective was to predict the IRS using a combination of physiological features obtained from wearable garment biosensors, making it a regression task. A Gradient Boosting algorithm [XGBoost (Chen and Guestrin, 2016)] was selected due to its ability to model complex relationships, handle missing data, and its robust performance (Chen and Guestrin, 2016). We now describe the steps taken in our machine learning model development pipeline in order to predict the inflammation response score. First, data were split into train and test sets while keeping the number of visits pre and post-inoculation balanced in both sets. Five randomly selected participants were kept in the test set, and the remaining participants (n = 50) were included in the train set (Figure 6A) which will be used to train and tune the ML models. Next, we applied filter methods to the train set to reduce features and improve model performance (Figure 6B). We removed duplicate features, features with a variance lower than 0.01 across the dataset (poor predictive potential since these features are nearly invariant across inflammation states), and highly correlated features (correlation threshold, r > 0.8 ). Finally, we employed hyperparameter tuning to optimize our machine learning model’s performance. Specifically, we used GridSearchCV from the Scikit-learn library to systematically search through a predefined hyperparameter grid that included “max_depth”, “learning_rate”, “n_estimators”, and “alpha”. GridSearchCV performed multi-dimensional tuning, exploring various combinations of these hyperparameters, rather than tuning one hyperparameter at a time. GridSearchCV essentially performed an exhaustive search, trying every possible combination of hyperparameters defined in our grid to identify the combination that resulted in the least mean squared error (MSE) of the predicted IRS. The grid search approach was implemented with a 3-fold cross-validation to reduce the risk of overfitting (Figure 6C). A 3-fold cross-validation when tuning the hyperparameters means that the train data is split into 3 approximately equal parts. The grid search algorithm iteratively uses each fold as a testing set while the remaining two folds are used as the training set. This process is repeated three times, with each fold serving as the testing set exactly once. Table 5 presents the hyperparameters, the ranges, the final tuned value and the step size during the tuning process.

We trained and tested models using a leave-one-subject-out strategy due to the limited number of data points (Figure 4). This approach involved training on the combined train data (n = 50 participants for training which corresponds to 153 visits) and the test participants’ data for 5 iterations (n = 5 participants for testing, 16 visits), excluding one participant during each iteration. Specifically, we use the 50 (train) + 4 (test) participants data to train the model and we test its performance on the left out participant from the test set and we repeat that with each test participant. To assess model performance, mean and standard deviation of performance metrics (across the 5 test participants) are reported. We developed 8 models to assess the predictive potential of different feature sources. For the BR, HR, MV signals, we developed 2 models based on features extracted from the 3 min of activity and the first minute of recovery, respectively. The model based on HRV features was only generated for the recovery period since HRV analysis is mainly relevant during rest/recovery (Boullosa et al., 2014; Kannankeril et al., 2004; Stanley et al., 2013; Goldberger et al., 2006). Additionally, we trained a model based on the features derived from the combined signals. We evaluated the model’s performance using the MSE values of the predicted IRS (Figure 6E), then, we derived a binary outcome from the predicted values by setting a threshold to 0, classifying participants as no inflammation (IRS ≤ 0 ) or high inflammation (IRS > 0 ) (Figure 5). This binary outcome enabled the evaluation of model performance via classification-related metrics (Figure 6G), including accuracy, sensitivity, specificity, precision, and AUC, which is the primary metric we used to evaluate the classification performance. The effect of varying thresholds in the range [-0.3, 0.3] is explored.

The XGBoost models were able to predict the IRS based on features derived from the wearable sensors (Table 6). Models trained on the activity period (3 min of stair activity) achieved a maximum AUC of 0.66, the highest one was trained on HR features only. Models trained on the first minute of recovery achieved AUC scores of 0.43, 0.60, 0.69 and 0.73 (95% confidence interval CI: 0.70–0.77) for the BR, HR, MV, and HRV feature sets respectively. The highest performing model (recovery, HRV, AUC = 0.73) achieved a sensitivity of 0.7 (95% CI: 0.66–0.74), a specificity of 0.77 (95% CI: 0.73–0.81), and a precision of 0.75.

Further details related to the model with the strongest performance (recovery period, HRV, AUC = 0.73) are now presented. The model comprises 8 features (see Table 7; Figure 7 for a description of each feature and relative feature importance, respectively). Feature importance is a measure that quantifies the contribution of each input feature to the predictive performance of the model. It helps us understand which features have the most significant influence on the model’s predictions. XGBoost measures feature importance by computing the gain achieved from each feature across all trees in the model. The gain represents the improvement in accuracy or reduction in loss resulting from a particular feature when used in a split decision within the trees. Features contributing to higher gain values are considered more important. In (Equation 5), the gain is calculated for each feature as it is used in each decision tree for splitting. The importance score for a feature is then determined by aggregating the gains across all trees, normalizing these values by the total sum of gains in the model.

For the best model, the binary threshold was varied over the interval [-0.3, 0.3] in steps of 0.1 which allows to obtain the ROC-AUC curve (Figure 8) (note that we cannot generate such curve automatically using Scikit Learn’s pre-existing methods since this is a regression task that was later converted to a classification task). Increasing thresholds led to maximizing specificity while decreasing thresholds maximized sensitivity (Table 8). Figure 8 is the corresponding ROC-AUC curve.

This study aimed to assess the potential of wearables for early detection of systemic inflammation following exposure to a low-grade VRTIs using machine learning algorithms and data collected during a controlled physical exercise (3-min CRSST). The model was able to predict the inflammation level using features derived from the wearable device. We were able to obtain the highest predictive ability using HRV features during the first minute of recovery after the 3-min CRSST.

The study design encompassed a structured timeline and a series of carefully planned visits and procedures. Even though moderate symptoms associated with viral replication are anticipated after the LAIV inoculation, we expected the participants’ immune systems to have a similar response as if exposed to mild influenza. The choice of the number of days pre-inoculation is justified by the fact that we wanted to establish a general baseline by capturing the most truthful trend within the participants’ daily lifestyle since a person’s physiological activity is not the same during all days of the week. Additionally, the choice of exactly 5 days post-inoculation is due to the fact that we wanted to capture all physiological states (no inflammation, high inflammation), especially the peak of inflammation, which is estimated to happen during the first 4 days after inoculation (McClain et al., 2015). Additionally, the use of the Astroskin/Hexoskin biosensor is backed up by its ability to ensure a non-invasive and convenient solution for collecting real-time data, allowing for a comprehensive assessment of participants’ respiratory and cardiac variables throughout the study.

In line with our second objective, we demonstrated that features derived from different data sources have different predictive abilities. The model utilizing BR features during the activity period exhibited limited performance. Similarly, the MV during the activity period also displayed relatively lower performance metrics. However, the model trained on HR features during the activity period performed better than the other features in the same period. In contrast, models focused on the recovery period based on BR and MV features showed improvement compared to their activity-based counterparts. Notably, the MV features during the recovery period exhibited higher AUC and specificity, suggesting its potential relevance in predicting systemic inflammation due to VRTIs during recovery periods. Moreover, the model trained on HRV during the first minute of recovery revealed highest sensitivity, precision, and AUC values among the analyzed features. Finally, the model using the combined features derived from the HR, BR, MV, and RRI signals demonstrated poor performance metrics. These results demonstrate that HRV features are particularly important in enhancing the model’s predictive performance. These findings are aligned with previous research (Boullosa et al., 2014; Kannankeril et al., 2004; Stanley et al., 2013; Goldberger et al., 2006), reporting that HRV may act as a physiomarker of cardiac wellness during post-exercise periods.

The superiority of recovery HRV over active HR and BR features is likely explained by several physiological mechanisms. First, the short duration of the 3-min CRSST is likely insufficient for participants to reach a steady-state metabolic condition, meaning that heart rate and breathing rate during activity primarily reflect transient acceleration kinetics rather than stabilized physiological responses. Second, while exercise induces a generalized sympathetic activation, the recovery period captures parasympathetic (vagal) reactivation and the restoration of autonomic homeostasis. Systemic inflammation is known to impair vagal reactivation, making post-exercise HRV a sensitive marker of inflammatory burden and autonomic resilience. Consequently, HRV during recovery acts as a functional autonomic “stress test”, providing greater discriminatory power than short-duration activity signals (Koeneman et al., 2021). Finally, heart rate outperformed breathing rate during activity because heart rate responds rapidly to neural vagal withdrawal, whereas breathing rate is driven by slower chemoreflex mechanisms related to carbon dioxide accumulation. As a result, the heart rate signal is more physiologically mature at the end of the activity period, although still less informative than post-exercise HRV.

Our study design allows setting a binary conversion threshold to the IRS. An additional exploratory objective was to assess the effect of varying the threshold level on model performance. This analysis confirms that the most balanced tradeoff between high sensitivity and specificity is attained when setting a threshold of 0. Other threshold values result in either higher sensitivity or specificity. This approach enables flexibility in adjusting the sensitivity and specificity of the model and it becomes particularly significant when considering the real-world implementation of our models in different scenarios and settings. The choice of the threshold should align with the specific objectives of the given context, taking into account the desired outcome and priorities.

Our results are not directly comparable to existing studies due to the uniqueness of our approach, which not only includes measurement of inflammatory biomarkers in blood samples, but also changes in cardiac and respiratory variables during a controlled physiological stress, i.e., 3-min CRSST. It must be regarded that variations in approaches, methodologies, features, and objectives between the studies have an impact and should be taken into consideration. Most studies relied on patient-reported diagnoses rather than quantitative measurements of inflammatory markers to predict VRTIs. In the previous iteration of our study (Hadid et al., 2025), we enrolled 55 healthy adults who continuously wore up to three different wearable devices (Oura ring, Astroskin–Hexoskin shirt, Biobeat watch), yielding more than 52,000 inflammatory biomarker datapoints from 1001 blood samples. Following inoculation, 78% of participants developed systemic inflammation, which allowed us to establish an objective ground truth for model development. The resulting dataset comprised 582 post-cleaning events, of which 18% were classified as inflamed based on an individualized coefficient of variance threshold. Physiological responses were observed as early as 12–24 h post-inoculation, and inflammatory surges peaked between 36 and 72 h. Using this rich biomarker-anchored dataset, we developed and tested eight wearable-based models. Our candidate model, relying on a minimal feature set from the Oura ring during sleep, achieved a ROC-AUC of 0.73 (95% CI 0.71–0.74) in real-time prediction and 0.89 (0.87–0.90) in the 24h-tolerance analysis (Hadid et al., 2025). In the context of VRTIs caused by COVID-19, Quer et al. (Quer et al., 2020) reported an AUC of 0.72 when relying on resting heart rate, sleep, and activity metrics. Mason et al. (Mason et al., 2022) and Conroy et al. (Conroy et al., 2022) achieved AUCs of 0.82 with sensitivity of 82% and 62% and specificity of 63% and 88%, respectively. Our model in this study shows slightly lower performance compared to these studies (AUC of 0.73 in real-time), but it is important to note that our results were achieved using objectively validated biomarkers rather than symptom reports (Hadid et al., 2025). Moreover, our classification of inflammation versus non-inflammation likely precedes symptom onset, suggesting a higher value for proactive monitoring in public-health and early-warning applications.

Several limitations are worth noting. The 3-min CRSST was performed under controlled laboratory conditions. The 3-min CRSST was also performed at a fixed rate of 30 steps/min. While the step rate is fixed/constant across all participants, the level of physiological stress will vary between between based on their age, sex, cardiorespiratory fitness and body mass. The CRSST is simple and requires minimal equipment (a single 20 cm step and a pacing device) making it feasible to perform in homes or workplaces with wearable monitoring; however, variations in step execution or cadence outside the lab could affect physiological responses Future studies should evaluate model performance in these real-world conditions. Potential participants could reproduce this task in real-world settings, but the quality of task execution may affect the performance of our algorithms. Ideally, strenuous physical activity (and recovery) would be automatically detected using the wearable devices and algorithms would be robust enough to perform under these types of uncontrolled activities. This makes further investigations imperative for validating these findings in diverse settings that account for representing real-world conditions. In addition, missing and noisy data led to a reduction in the dataset size. The utilization of consumer-grade wearable devices (smaller and more convenient for every-day use compared to the Astroskin that comes with a head band and requires the user to put a relatively large device measuring the acceleration in the correct orientation) for data collection, while not eliminating the possibility of recording missing and noisy data, would allow for more extensive data collection to be conducted. Furthermore, the group of participants in our study was limited since we specifically focused on healthy adults. It is important to recognize that extending our conclusions to groups like the elderly or vulnerable populations may require additional investigation. These groups may exhibit variations in their physiological responses to infections, exercise, and unique profiles of biomarker changes.

Additionally, It is important to note that individual differences in physical fitness and lifestyle factors, such as caffeine consumption, may influence physiological responses during the stair-stepping test. Further, participants with significant mobility limitations were excluded, and therefore our findings are not be generalizable to individuals who are unable to perform such exercise tasks. Future studies could incorporate stratification by fitness level, standardized dietary controls, and alternative exercise or passive monitoring protocols to expand applicability. Consequently, future research should aim to replicate and adapt our methodology to these diverse populations, ensuring broader applicability and effectiveness of our early detection approach. Finally, it is essential to recognize the limitations of our study concerning the specificity of detectable inflammations. Our emphasis on respiratory inflammation and specific biomarkers associated with this context may not encompass the full spectrum of inflammatory responses within the body. Inflammatory processes can vary significantly depending on underlying causes and affected systems. While our model displays promise in detecting systemic inflammatory surges due to viral respiratory tract infections, its specificity may be more confined to this particular inflammatory context. Researchers should exercise caution when applying our findings to broader inflammatory conditions, considering the need for context-specific models in such cases.

Several strategies could be employed to address the limitations identified in this study. Future work could stratify participants by fitness level, standardize dietary and caffeine intake, and record environmental or seasonal factors. Alternative exercise protocols or passive monitoring could enable inclusion of participants with mobility limitations.