Table 2.

Classification of data analytics for wearable-generated temperature data.

Types of analytics and description of the included studies			Temperature features		Analysis techniques
Descriptive analytics (n=19)
	Compared differences in temperatures recorded across monitoring methods, patient cohorts, or anatomical sites [43,44,47]	Mean, SD, median, and maximum		Student t test and ANOVA
	Counted the numbers and frequencies of febrile events and compared the time difference between fever episodes detected by continuous and intermittent monitoring [17,38,39,42,47,48,50,52,58,61]	Timing of fever		Student t test
	Developed a heuristic scoring system to quantify the clinical relevance of fever detection, incorporating fever duration (short spike <60 minutes; prolonged episode ≥60 minutes) and clinical actions triggered [53]	Thresholds and fever duration		Descriptive comparison
	Hourly points from the continuous temperature data were used to simulate 1 to 24 intermittent daily measurements to identify the optimal timing and frequency achieving the highest fever detection rate [60]	Counts and frequencies		Brute-force strategy
	Compared the performance of fixed, individualized (baseline mean + 2SD), and combined thresholds for axillary temperature in detecting cytokine release syndrome [59]	Thresholds		Grid search
	Detected abnormal values or temperature deviations from preset thresholds to trigger alerts [17,61]	Thresholds		Described the counts and frequencies
	Visualized patterns of temperature, such as rising trends or fluctuations in temperature curves [41,46,51]	Visual patterns		Line plots
	Described the disappearance of the circadian rhythm of skin temperature in patients with critical illness, in comparison with recovered patients [49]	Circadian rhythmicity		Cosinor models
Diagnostic analytics (n=6)
	Classified fevers of infectious or non-infectious causes using temperature trajectory patterns within 4 hours around fever start time [37]	Trajectory patterns		Dynamic time warping and k-means clustering
	Quantified fever intensity and evaluated its association with epidural analgesia duration and dosage to identify risk factors for fever [62]	Area under time-temperature curve		Logistic regression and ROCa curve analysis
	Examined the correlation between skin temperature and oral/ear canal temperature and identified a diagnostic cut-off value of skin temperature for detecting fever or infections [44,48,57]	Correlations		ROC curve analysis
	Analyzed the correlation between standard deviation and peak temperature to reveal an increased temperature variability before and after the spike [39]	Correlations		Pearson correlation and linear regression
Predictive analytics (n=9)
	Predicted SCCb by inputting temperature data into a deep neural network to calculate an SCC score reflecting dissimilarity from normal patterns [45]	Cosine similarity		Deep neural network
	Used aggregated temperature features (median, variance, coefficient of variation) as inputs for pulmonary infection prediction models [56]	Median, variance, coefficient of variation		XGBoostc
	Extracted residual deviations of temperature from the patient's individual circadian rhythm pattern as early signs of febrile adverse events [6]	Circadian residuals		Circadian modeling and genetic algorithm
	Used continuous temperature data as inputs for machine learning models to predict the risk of advanced sepsis [54], cytokine release syndrome [59], or COVID-19 deterioration within the next 24 hours [63]	Mean, SD, extreme value, and percentage of time spent over thresholds		Machine learning models
	Used skin temperature features (eg, mean and maximum) within various time windows before fever onset as early predictive signals for impending fever episodes [40,55,57]	Correlations		Logistic regression and generalized estimating equation

^aROC: receiver operating characteristic.

^bSCC: serious clinical complications.

^cXGBoost: extreme gradient boosting.