Data Collection.

The original dataset joins individual CDRs that MNOs routinely gather for billing purposes with individual-level ILI diagnosis data from Iceland’s Centre for Health Security and Communicable Disease Control (CHS-CDC), which collects and stores all records of ILI diagnoses in Iceland. We developed and used a privacy-preserving data hand-off and merging protocol approved by Iceland’s Bioethics Committee (Vísindasianefnd): A large MNO sent encrypted phone identifiers (IDs) and national ID numbers (NINs, which are public information in Iceland) to the CHS-CDC. The CHS-CDC supplied dates of ILI diagnoses for NINs and then replaced NINs with an anonymous encrypted identifier before providing the data to us (SI Appendix, Data Linking and Privacy). The MNO provided us with CDR data (SI Appendix, Mobile Network Data) containing the encrypted IDs of the phones on either side of a call, the timestamp, the length of the call (in seconds), and the geographical coordinates of the cell-phone towers that interacted with the phones (SI Appendix, Table S1). No demographic or private data, such as age, gender, or contents of calls or texts, were included. The cell tower accessed during normal phone use provides a proxy for the device’s location. The granularity of location varies with locality—regional tower density increases proportionally with regional population (Fig. 1). At the time, MNOs provided cell coverage for virtually all residences in Iceland, either directly through their network or through a roaming service. We filtered out individuals with multiple subscriptions (SI Appendix, Data Preprocessing). Using phone-ownership information, each phone was matched to the DoD of its owner for the subset of users that pay only for one phone. This postprocessed subset, referred to as the dataset below, accounted for 25 to 30% of the MNO’s users and encompassed all data analyzed in our paper. We defined the home tower of each individual as the tower that picks up more calls and texts between midnight and 8 a.m. than other towers. The distribution of home-tower locations was strongly correlated with residential census counts for the corresponding postal codes for our dataset ($$ r=0.86$$, $$ P<8\times 10^{-49}$$) and among those with ILI diagnosis ($$ r=0.88$$, $$ P<2\times 10^{-43}$$). The home towers were thus spatially representative for the entire Icelandic population. We focused our analysis on the 1,434 diagnosed users who generated sufficient CDR data to establish a home-tower location in the 4-wk period centered on their DoD.

Feature Extraction.

To characterize user behavior, we extracted 36 features (independent variables) from both incoming and outgoing CDR data encompassing movement, activity, and social-network behavioral patterns (SI Appendix, Feature Extraction). Most features exhibited a right-skewed distribution (SI Appendix, Fig. S2) and shared general characteristics across control and diagnosis groups. They include the following (boldface in Table 1).

Table 1.

Feature characteristics from the 29-d period around each individual’s DoD (additional characteristics are in SI Appendix, Table S2)

	Control	Diagnosed
Feature	Mean	Mean	Anomalous days
Number of towers visited	3.04	2.74	−12, −2, −1, 0, 1, 2, 3, 4
Number of new locations visited	0.5	0.43	1, 3
Unique contact count
Incoming	2.25	2.02	$$ \varnothing $$
Outgoing	2.50	2.28	1, 2, 3, 6
Both	4.04	3.67	2
New contacts
Incoming	0.61	0.50	$$ \varnothing $$
Outgoing	0.65	0.55	1, 3, 6
Both	1.19	1	0, 1, 3
Call duration, total, s
Incoming	190	480.51	0, 1, 2
Outgoing	162	435.09	0
Both	479.5	915.60	$$ \varnothing $$
Calls count
Incoming	3.10	2.84	$$ \varnothing $$
Outgoing	3.60	3.37	1, 2
Both	6.66	6.22	1
Texts count
Incoming	2.71	2.87	−10
Outgoing	1.74	1.93	1, 2
Both	4.46	4.79	−10, 11
Calls and texts count
Incoming	5.78	5.71	$$ \varnothing $$
Outgoing	5.34	5.30	1, 2
Both	11.12	11.01	2
Mean call duration
Incoming	133.35	140.74	−1, 0, 1, 2, 4, 11
Outgoing	107.96	106.74	$$ \varnothing $$
Both	143.14	149.84	−8, −1, 0, 1, 2, 3, 4, 11, 13, 14
Top 3 contacts by duration
(SI Appendix, Sensitivity Analysis)
Incoming	0.68	0.69	$$ \varnothing $$
Outgoing	0.70	0.69	$$ \varnothing $$
Both	1.37	1.38	1
Remaining contacts by duration
Incoming	1.57	1.33	$$ \varnothing $$
Outgoing	1.80	1.58	1, 2, 4, 6
Both	3.37	2.92	1, 4, 6
Top 3 contacts by frequency
Incoming	0.58	0.56	$$ \varnothing $$
Outgoing	0.63	0.61	1, 2
Both	1.21	1.17	1
Remaining contacts by frequency
Incoming	1.67	1.46	0
Outgoing	1.88	1.67	−9, 2, 6
Both	3.54	3.13	$$ \varnothing $$

Number of towers visited measures the number of unique tower coordinates connected to by the cell phone within a time interval (bin). This feature helps describe movement during the time period, but can inflate in areas where multiple towers can provide cellular signal.

Mean call duration (incoming and outgoing) measures call activity by dividing the total duration of calls by the number of calls the user placed or received in the time interval.

Number of calls (outgoing) measures the number of calls placed by the device in the time interval.

Departure from Routine Behavior.

We use $$ x_{\mathit{fi}d}=E_f\left(i,d\right)$$ to denote the raw feature value for a feature $$ f$$, extracted from the CDR by function $$ E$$, for individual $$ i$$, and on day $$ d$$. Extraction is performed for all features $$ f$$ in Table 1.

To control for the weekly behavioral routine of individual $$ i$$, each feature value is detrended through linear regression over values of the same weekday in the past $$ W$$ weeks. Specifically, let

We used $$ W=10$$ weeks of past data to correct for seasonality in our experiments, which gave comparable results to an alternative approach to detrending based on ranking features and normalizing them (SI Appendix, Seasonality).

Based on the data, we used a linear model to capture the change in values over time $$ p_j=\beta j+\alpha +{\epsilon }_j$$ with errors $$ {\epsilon }_j$$ for each $$ j\in J$$; we fit parameter values for $$ \widehat{\alpha }$$ and $$ \widehat{\beta }$$ to minimize the squares regression error

Control Group.

Each diagnosed individual was matched with a control individual from the undiagnosed group, based on home location. All measurements thus far have applied to individuals diagnosed with ILI during the epidemic. To compare the diagnosed population against a control population, a subset was selected from the rest of the data—those not diagnosed for ILI were assumed to be uninfected, though they may show behavior consistent with symptoms but are well, or have ILI symptoms but did not use health services. Of 74,644 people, we were able to identify home towers for 36,140. Each diagnosed person’s control was selected randomly from the undiagnosed individuals among the 36,140 who shared a home tower with the diagnosed individual. For this dataset, control selection exhibited no noticeable differences across three methods: selecting randomly, matching for home tower, or matching home tower and frequency of calls (36).

We analyzed the pattern differences between the means of the detrended feature values ($$ z_{\mathit{fi}d}$$) of the individuals in the two groups. The 29-d range (2 wk either side of DoD) centered around every diagnosed individual’s DoD range $$ \left[-\mathrm{14,14}\right]$$, with DoD mapping to zero. Controls used the same days of data as their diagnosed match. The average deviation from weekly routine on all days in the range was compared (SI Appendix, Fig. S9) with original feature values ($$ x_{\mathit{fi}d}$$), shown in SI Appendix, Figs. S2, S3, and S8.

Statistical Comparison.

We compared the behavior of the diagnosed and control groups across each detrended feature value $$ z_{\mathit{fi}d}$$ and each day using the Kolmogorov–Smirnov (KS) statistic. To counteract the increase in type I errors caused by running multiple significance tests, we used the Benjamini–Hochberg (BH) procedure to control the false discovery rate (FDR), as it presents the most conservative FDR correction for this mix; the adjusted P values can then be used to assess the evidence for or against the null hypothesis. The BH procedure assumes independent tests. Some tests act on dependent, interacting samples—e.g., a value on a specific day is ranked against values from the same day of the week for several weeks prior—whereas others are independent tests. Confidence bands for the KS test were computed and plotted for each day of the primary three features deemed significant based on the P values with the FDR correction (SI Appendix, Fig. S9). The significance test and the CI calculations use $$ \alpha =0.05$$.

Less Movement.

The number-of-towers feature indicates that the diagnosed group tends to travel less than usual, even before diagnosis. Such lower travel patterns coincide with the typical symptomatic period of influenza (37). The maximum effect is observed on the day following diagnosis, when diagnosed individuals travel to 1.1 to 1.4 fewer locations than normal. Differences are observed between the diagnosed and control groups from 2 d prior to the DoD until 4 d after DoD ($$ KS>0.084$$, $$ P<3.2\times 10^{-3}$$; Fig. 2 and SI Appendix, Fig. S10). Other days in the 4-wk period display the diagnosed and control groups acting similarly.

Fig. 2.

Changes in average phone-use behavior associated with diagnosis. (Left) Users visit fewer locations on days around diagnosis. (Center) They make and receive longer phone calls on days near diagnosis. (Right) They initiate fewer calls on the days after diagnosis, with the exception of the day of diagnosis itself. Graphs display the mean deviation from “normal” routine behavior ($$ z_{\mathit{fi}d}$$) for each group on the relative day of illness determined by DoD (day 0). CIs (2.5 to 97.5%) are calculated using bootstrapping (SI Appendix, Visualization).

Longer Calls.

Mean call duration shows that people tend to make longer calls on average on the day after the DoD (Fig. 2), when significant differences are observed between the diagnosed and control groups ($$ KS=0.155$$, $$ P<4.6\times 10^{-10}$$; SI Appendix, Fig. S22). On the day following diagnosis, diagnosed individuals spend an average of 41 to 66 s longer on the phone than usual.

Fewer Calls Placed.

Number of outgoing calls gives another perspective of behavior following diagnosis. Although call duration increases around DoD, the number of outgoing calls decreases on the day after the DoD, with an average of 2.3 to 3.3 fewer calls than is routine ($$ KS=0.102$$, $$ P=5.6\times 10^{-4}$$; SI Appendix, Fig. S18). On the day of diagnosis, diagnosed individuals increase outgoing calls relative to their routine compared to the days before and after.

Statistical significance through FDR-corrected P values is supported by KS CIs for nearly all comparisons (Fig. 2 and SI Appendix, Fig. S9). Notably, the diagnosed group displays significant changes in mobility, even prior to seeking healthcare and receiving a diagnosis (SI Appendix, Visualization).

Limitations.

The results depend on the metadata arising from mobile-phone use, presenting both advantages and drawbacks (21, 33). The increased data bandwidth provided by MNOs and rapid device and app development over the past decade have altered user behavior patterns to communicate more via internet-based applications and less via calls and text. In our dataset, cellular internet data access (denoted general packet radio service [GPRS]) provided additional location information to CDR records of calls and texts, a situation that has likely shifted since the H1N1 outbreak (SI Appendix, Comparing CDR and GPRS Data). At 3 y following the epidemic, the Icelandic CDR and GPRS data contained a stronger location proxy than in 2009 due to more smartphone apps periodically connecting to cellular towers for Internet access, but poorer information for features pertaining to call duration, frequency, and top contacts.

Since many nations experience limited Internet access [53.6% of the world population in 2019 (22)] and smartphone availability [39.4% worldwide (38)], it would be reasonable to assume that call and text usage in these locations may follow similar patterns as in our dataset, but we caution against assuming all cell-phone behavior to be universal (33). Further, mobile-phone ownership may bias against those in greatest need of public health intervention. The results report aggregate behavior changes, which are likely to include patterns caused by other illnesses or injuries. Our approach depends on maintaining individual-level behavioral histories, since the signal we identified concerns departure from routine behavior rather than the actual behavior itself, as seen by comparing the raw and detrended distributions 6 d prior to diagnosis (SI Appendix, Fig. S2) with the day following the DoD (SI Appendix, Fig. S3). Finally, Iceland contains a small, mostly homogeneous, and generally affluent population bound to an island, with idiosyncratic behavior, including unusually high mobile-phone usage. Seasonal effects may be exaggerated in Icelanders compared to other populations due to Iceland’s proximity to the Arctic.