High frequency temperature monitoring for early detection of febrile adverse events in patients with cancer

Fever is an important early sign of serious treatment-related adverse events, such as cytokine release syndrome (CRS) caused by chimeric antigen receptor T-cell (CAR-T) immunotherapy, and infection related to chemotherapy-induced neutropenia (Oved, Barrett and Teachey, 2019), commonly experienced by patients with cancer. The standard approach for detecting fever in hospitalized patients is intermittent temperature monitoring, typically every 4-8 hours, which could lead to inherent delays in diagnosis of febrile adverse events.

The availability of non-invasive, wireless, wearable sensors to “continuously” monitor body temperature raises the possibility of earlier detection and diagnosis of fever and its associated adverse events. Some studies have begun to investigate this possibility (Jordan et al., 2017; Sampson et al., 2019; Liu et al., 2020; Smarr et al., 2020), yet a systematic investigation in patients with cancer using FDA-approved devices and comparing to standard-of-care (SOC) monitoring is needed. Furthermore, the large volume of high-frequency temperature data that can be obtained from a wearable sensor opens the possibility of carrying out computational analysis to identify signals for anticipating fever before it occurs.

To investigate these possibilities, we conducted a prospective study in 68 patients receiving hematopoietic stem cell transplant (HCT) or CAR-T therapy in the inpatient setting (Figure S1A). After providing IRB-approved informed consent, patients were asked to wear a self-administered, non-invasive, and FDA-approved wearable sensor (TempTraq®, BlueSpark Technologies), applied as an axillary skin patch according to manufacturer’s instructions, to capture high-frequency temperature measurements (HFTM) every 2-minutes; data were wirelessly transmitted in real-time to a cloud-based server (Sampson et al., 2019). HFTM data from 62 patients (n=39 HCT, n=23 CAR-T) were available for analysis to compare timing of fever detection with SOC temperature measurements, which are typically taken every 4-8 hours by nursing staff as part of routine clinical care. During the monitoring period, we collected a total of 585 days of HFTM data across all 62 participants with a median data capture of 8.5 days/patient. When patients were wearing an HFTM patch, we collected ~90-fold more data points with HFTM (n = 421,367) than SOC (n = 4,816).

We first evaluated the timing of SOC- and HFTM-detected fevers and found that HFTM detected 89% (24/27) of these fevers a median 5.5 hours (h) earlier than SOC (Figure S1B). For three fevers detected earlier by SOC, the median time was 1.9 h earlier. Overall (n=27 fever events analyzed), HFTM showed a median 4.9 h earlier detection time than SOC.

As expected, most fevers detected in patients having received CAR-T therapy were related to CRS; whereas in HCT patients, infection-related fevers were more common (Figure S1C). Interestingly, we found that fevers caused by infections were detected by HFTM significantly earlier (median = 18.5 h) than by CRS (median = 4.4 h; p = 0.012, two-tailed t-test); examples are shown in Figures S1D–F.

We further investigated whether we could computationally identify potentially predictive signals that precede fever (i.e., before an HFTM-detected temperature rise to 38°C). We hypothesized that subtle perturbations in temperature dynamics may be discernible prior to fever and may manifest in circadian modeling analysis as deviations from baseline circadian pattern. To test our hypothesis, we fit a circadian profile based on 24-h of preceding data for every data point leading up to independent HFTM fever events that had sufficient data for circadian modeling (Figure S1G). This approach allowed for real-time updating of the circadian profile (magenta curves) with an average temperature measurement (green lines), and amplitude and phase (the time of the minimum of the circadian profile) estimates.

From the circadian fit, we computed circadian residuals (Figure S1H, blue dots), defined as the difference between the circadian fit and the data point recorded. To incorporate changes in average temperature that may have occurred across larger time scales, we also computed a standardized residual of the average temperature (orange dots). Finally, for our subsequent analysis, we computed a total residual, defined as the sum of the circadian and average temperature residuals (red dots). We separated the total residuals into two groups based on periods of time: pre-fever day residuals (i.e., residuals from 24-h immediately prior to fever, red shaded region) and a patient-specific baseline (i.e., residuals calculated prior to the pre-fever day, blue shaded region).

We then projected the patient-specific baseline residuals onto the pre-fever day, matching the phase estimates for each data point to the phase of the pre-fever day to account for specific measurement bias at certain times of the circadian cycle. We computed the residual difference (Figure S1I) between the pre-fever day and the patient-specific baseline period and predicted 95% confidence intervals after sampling with replacement 1000 times. In general, the residuals in the pre-fever day began to deviate positively from those of the patient-specific baseline ~12-h prior to fever. Moreover, this deviation was sustained and statistically significant from ~3.5-h pre-fever up to fever onset, with additional discrete spikes of statistical significance ~5-6 and 8-h before fever onset, collectively demonstrating signals in HFTM data that presaged the occurrence of fever.

Taken together, our results demonstrated the potential of an HFTM approach using wearable sensors to provide considerable lead time (4.9 h earlier than SOC) for early detection of febrile adverse events, with the potential to add 3.5 h or more lead time by circadian modeling. This duration of lead time is clinically significant for patients with cancer who are commonly immunocompromised and at risk for infection, because time-to-first antibiotics can play an important role in subsequent mortality in neutropenic fevers and sepsis (Mullen et al., 2000; Wingard, Hsu and Hiemenz, 2011), especially in the setting of septic shock where mortality increases with every hour of delay in antibiotic administration (Kumar et al., 2006). Our inpatient study provides a foundation for investigation in the outpatient setting. In particular, the impact of early detection of infection may be seen at a larger scale amongst outpatients with cancer receiving chemotherapy who are at-risk for febrile neutropenia. Our data was collected using an FDA-approved wearable sensor suitable for home use, making it readily implementable.

Furthermore, the lead time provided by HFTM is also clinically relevant to monitoring patients treated with CAR-T. It can enable earlier intervention in CRS through escalation of care, including earlier administration of anti-cytokine therapies [e.g., tocilizumab (an IL-6R antagonist)] (Oved, Barrett and Teachey, 2019), which may reduce life-threatening morbidity and mortality associated with CRS. This lead time could also facilitate the transition of extremely expensive inpatient CAR-T care to the outpatient setting, since the lead time could provide sufficient time to return to hospital in case of impending CRS.

We hope that our results will spur more in-depth investigation of the HFTM approach in patients with cancer, ultimately through prospective clinical trials. Elements to be investigated in future work include: optimizing patient education and support to minimize missing data; developing computational algorithms to probabilistically identify the cause and clinical actionability of a fever from temperature dynamics and additional clinical data; and increasingly individualizing prediction and detection of febrile adverse events using a patient’s own baseline temperature pattern as a reference, rather than a one-size-fits-all 38°C threshold approach.

Supplementary Material

Figure S1

Figure S1. High frequency temperature monitoring using a wearable device and computational analysis for early detection and anticipation of febrile adverse events in patients with cancer.

A. Overview of study design. We recruited 68 participants for continuous temperature monitoring who received either chimeric antigen receptor T-cell (CAR-T, 25/68) therapy or hematopoietic cell transplant (HCT, 43/68) for the treatment of hematological diseases. Six participants were excluded either due to a lack of inpatient monitoring (2/68) or insufficient data collection (4/68 had a <1 day monitoring period). For all participants, we collected high-frequency temperature monitoring (HFTM) data using a TempTraq® wireless, non-invasive axillary skin patch sensor, clinical standard-of-care (SOC) temperatures from the electronic health record (EHR), as well as annotation of clinical data related to febrile events using the EHR. Presence or absence of cytokine release syndrome (CRS) for each day of hospitalization was annotated using standard clinical criteria (Lee et al., 2019). Annotation of the cause of febrile events into categories of infection, CRS or other cause was based on retrospective EHR review, including review of daily clinical team assessment notes and laboratory studies (e.g., positive cultures or other diagnostic tests for infection). B. HFTM enables early detection of clinical SOC fever. To define independent SOC fever events, we used a common clinical definition within the CAR T-cell therapy and HCT settings: a temperature ≥ 38°C with no fever recorded in the past 24 hours (h) (Freifeld et al., 2011). To define independent fever events for HFTM data, in which data points are more abundant, we added to the SOC fever definition a requirement that any HFTM-detected fever contains at least 3 recordings ≥ 38°C within the first hour of fever, and that the fever duration must be greater than 1 h. Fever duration was defined as the time duration between the first and last temperature recordings ≥ 38 °C such that the last recording had no temperature recording ≥ 38°C in the following 24 h. To evaluate HFTM for the early detection of fever, HFTM early detection time was calculated as the difference of the time of HFTM fever detection relative to the start of the SOC fever. We focused on 27 clinical fevers with sufficient HFTM data coverage that enabled comparison with SOC (i.e., >= 20% HFTM data coverage in the hour preceding SOC fever onset). HFTM early detection time was plotted for each clinical fever and was stratified by fever cause (CRS = red, Infection = green, Other Cause = blue). Infections included a range of bacterial, viral and fungal causes as well as two severe mucositis cases, including one treated empirically with antibiotics by the clinical team. Negative values represent earlier detection by HFTM (n = 24; median early detection time = −5.5 h), while positive values represent earlier detection by SOC (n = 3, 1.9 h). Taking into account all fever events shown here (n = 27), HFTM showed a median early detection time of −4.9 h.C. HFTM shows differences in early detection time between different fever causes. Early detection time, stratified by fever cause, was plotted for all fevers detected early by HFTM (Other Cause: n = 6, median early detection time = −4.8 h; Infection: n = 8, −18.5 h; CRS: n = 10, −4.4 h). Boxplots represent 95 % confidence interval (CI) with interquartile range and median values. D-F. Examples of two common temperature patterns of HFTM early detection of fever. D shows one pattern that we observed, illustrated by HFTM (blue) and SOC (orange) around the first clinical fever of patient CART-01, which was caused by CRS. Here, HFTM detected a sustained temperature spike until fever detection by SOC. E shows a more extreme case of this first pattern, where data collected around the first clinical fever of patient HCT-27, which was caused by an infection, showed an HFTM early detection time of nearly 27 h. F shows a second temperature pattern of HFTM early detection, illustrated by data collected around the first clinical fever of patient HCT-05, which was caused by infection. In this case, HFTM detected a transient episode of fever nearly 24 h before the subsequent SOC fever became apparent. G-I. Circadian-informed modeling of HFTM data reveals statistically significant deviations from baseline prior to fever. As an example to illustrate the method, temperature data over several days leading to fever onset (0 h) from patient CART-08 is shown in panel G. For each data point after the first 24 h, we fit a circadian profile (magenta) to the preceding 24 h of data characterized by the function

$$C(t)=T_m+Acos(\pi /12(t-\phi )),$$

where t is the time in h, T_m is the average temperature in the 24 h (green line), A is the amplitude of the circadian profile, and φ is the phase estimate. We fit the three parameters T_m, A, and φ using the genetic algorithm, a global optimization procedure in MATLAB 2020RB, to minimize the L₂-norm difference between the data (t_i, T_i) and the fit estimate (t_i, C(t_i)). We computed the circadian residual (magenta dots) as

$$R_C(t_i)=T_i-C(t_i),$$

where T_i is the temperature at t_i. That is, the circadian residual is the difference between the fit and the data point. The residual was positive if the data point was greater than the value from the fit, and conversely the residual was negative if the data point was less than the value from the fit. For data points after the initial 24 h, t_i, we fit the three parameters of C(t) to all data between t_i - 24 and t_i. Then, we used the fit C(t) on that 24 h of data to compute the residual R_C(t_i) as above. We continued this process until we computed a circadian residual for every data point leading up to the fever. H. The blue dots are the standardized circadian residuals from panel G, and the orange dots are the standardized average temperature residuals. We computed a total residual for subsequent analysis as the sum of the circadian and average temperature residuals (red dots):

$$R_T(t_i)=\frac{(R_C(t_i)-mean(R_C(t_j)))}{std(R_C(t_j))}+\frac{{{(T}^i}_m-mean({T^j}_m))}{std({T^j}_m)}.$$

Here, j = 1,…,i−1 and T_m^j corresponds to the mean temperature estimate at time t_j. We then divided all data before a fever into two subsets: data in the 24 h leading up to the fever (pre-fever day, red-shaded region) and all data before the pre-fever day (patient-specific baseline, blue-shaded region). I. For fever events that had sufficient days of HFTM data available to enable the analyses described above (n = 17), we computed the difference between the total residuals from the pre-fever day and patient-specific baseline (calculated as described in G). For each data point (t_i, R_T(t_i)), we took t_i = t_i modulo 24. This translates the time point t_i to a time between 0 and 24. Then, we computed the phase difference

$${\phi }_{diff}={\phi }_f-{\phi }_i,$$

where ${\phi }_f$ is the phase estimate at fever onset and ${\phi }_i$ is the phase estimate at t_i. After, we shifted the time point t_i in the following way

$${\overline{t}}_i=(t_i+{\phi }_{diff})\mathrm{mod}24.$$

This shifts the phase of the time point t_i to match the phase at fever onset. Finally, we generated the patient-specific baseline data set by taking the mean of (${\overline{t}}_i,R_T(t_i))$ in two-minute bins from 0 to 24 hours to match the time resolution of the fever data sets.

We ran the residual comparison on 17 fever and patient-specific baseline data sets. First, we fit a nonlinear profile to both the 24-h patient-specific baseline residuals and the 24-h pre-fever day residuals for each participant. We used the bs function in R to fit a b-spline with 22 internal knots and 2 boundary knots. From the nonlinear fits, we estimated the total residual at 15-minute intervals. Then, we took the difference between the fever residual estimate on the pre-fever day and the corresponding residual estimate from the 24-h patient-specific baseline estimate. Finally, we took the mean difference of the 17 fever events every fifteen minutes. We repeated this process 1000 times after sampling with replacement from the 17 fever events to generate a distribution of mean differences between the patient-specific baseline period and the pre-fever day every fifteen minutes from 24 to 0 h before fever onset. From the null distribution, we used the quantile function in R to compute a 95% quantile of the data with a probability range from 0.025 to 97.5. Then, any time point with a mean difference outside of this quantile was considered statistically significant (see I, shaded regions) with p-value <0.05.