Concordance between GPS-based smartphone app for continuous location tracking and mother’s recall of care-seeking for child illness in India

Background Traditionally, health care-seeking behaviour for child illness is assessed through population-based national demographic and health surveys. GPS-based technologies are increasingly used in human behavioural research including tracking human mobility and spatial behaviour. This paper assesses how well a care-seeking event to a health care facility for child illness, as recalled by the mother in a survey setting using questions sourced from Demographic and Health Surveys, concurs with one that is identified by TrackCare, a GPS-based location-aware smartphone application. Methods Mothers residing in the Vadu HDSS area in Pune district, India having at least one young child were randomly assigned to receive a GPS-enabled smartphone with a pre-installed TrackCare app configured to record the device location data at one-minute intervals over a 6-month period. Spatio-temporal parameters were derived from the location data and used to detect a care-seeking event to any of the health care facilities in the area. Mothers were asked to recall a child illness and if, where and when care was sought, using a questionnaire during monthly visits over a 6-month period. Concordance between the mother’s recall and the Track- Care app to identify a care-seeking event was estimated according to percent positive agreement. Results Mean concordance for a care-seeking event between the two methods (mother’s recall and TrackCare location data) ranged up to 45%, was significantly higher (P-value < 0.001) for care-seeking at a hospital as compared to a clinic and for a health care facility in the private sector compared to that in the public sector. Overall, the proportion of disagreement for a care-seeking event not detected by TrackCare but reported by mother ranged up to 77% and was significantly higher (P-value < 0.001) compared to those not reported by mother but detected by TrackCare. Conclusions Given the uncertainty and limitations in use of continuous location tracking data in a field setting and the complexity of classifying human activity patterns, additional research is needed before continuous location tracking can serve as a gold standard substitute for other methods to determine health care-seeking behaviour. Future performance may be improved by incorporating other smartphone-based sensors, such as Wi- Fi and Bluetooth, to obtain more precise location estimates in areas where GPS signal is weakest.

The CONSORT statement for randomized trials of non-pharmacologic treatments was used to guide the process and reporting of this study [40]. The Improving Coverage Measurement for Maternal, Newborn and Child Health Study was conducted in 22 villages in Pune district in Maharashtra State in the western region of India. The study area comprises a population of about 143 000, is situated about 30 km from Pune city metropolis with rapid urbanization and industrialization [41]. Two rural hospitals, several health centres and more than a hundred private medical practitioners supported by 68 chemists provide medical care. The majority (71%) of these providers are situated in four villages along the main State highway that passes through the study area (Figure 1). Map of study area showing location of HCF and study participants, Vadu, India. Map zoomed in the inset to show the high density of health care facilities in one of the villages of the study area. A total of 926 mothers aged 15 – 49 years having at least one living child under the age of five years were randomly sampled from a population database maintained by the Vadu Health and Demographic Surveillance System. The mothers who were sampled were further randomly assigned to one of three groups – longitudinal phone group (200 mothers given a smartphone with a TrackCare App who were followed up monthly for 6 months), longitudinal control group (100 mothers who were followed up monthly for 6 months), and six cross-sectional control groups (about 75 mothers in each group who were followed up on one occasion during the six month period) (Figure 2). Field workers approached mothers at home for recruitment, 749 of whom (response rate 81%) consented to participate in the study. The investigators were not blinded to the mother’s group assignment. The study included a longitudinal control group to adjust for the potential bias in reporting care-seeking due to the presence of the study phone. The cross-sectional control groups were included to determine whether changes in care-seeking reports were due to repeated administration of the survey questionnaire. Sample size for the phone group was calculated based on an estimated 15-day care-seeking prevalence of 20% [42], an average of two eligible children per enrolled mother, a base concordance of accurate care-seeking of 80%, and a precision level of 8%. Participant follow-up schedule. Cross-sec. comp. group – Cross-sectional Comparison Group; Long. comp. group – Longitudinal comparison group. Baseline information collected from mothers included individual and household demographic and socio-economic characteristics, and care-seeking preferences for child illness (Appendix S1 in Online Supplementary Document(Online Supplementary Document)). Mothers in the phone group were asked to rate their potential concerns (about damage, theft, loss of phone, personal safety, etc.) regarding the use of smartphone. These questions were developed based on our experiences with devices used for direct electronic data capture in the Vadu Health and Demographic Surveillance and other studies [43]. At each follow up, questions identical to those in the NFHS questionnaire were used to ask the mother to recall if the child had diarrhoea, fever or cough within the last 15 days, if care was sought, and the type of provider. Additional questions were asked to find when (how many days before the follow up) and where (name of health care facility) the care was sought. Mothers in phone group were further asked whether the study smartphone was carried during the visit to the health care provider. A separate health care facility survey identified and collected geo-coordinates of all fixed health care facilities (hospitals, clinics, health centres, chemist shops etc.) in the area. Mothers in the phone group received a dual SIM, GPS-enabled Sony Xperia E4 smartphone with one SIM card and the pre-installed TrackCare app. To encourage its use by the mother, all costs towards voice call, instant messaging and data usage was paid for by the study. Mothers could additionally opt to install their personal SIM card in the secondary SIM slot to ensure continuity in use of their existing SIM card and mobile number. Mothers were trained by the field investigator during home visit in the key features of the smartphone, instructed not to change the phone location access settings (set to high accuracy – ie, the location data sourced and retained from the better of the two sources viz. the network provider and the GPS chipset itself), ensure that the phone battery was kept charged during the whole day, and encouraged to carry and use the smartphone when moving outside of home. The smartphone settings were checked, any deviations from optimal settings were corrected, and the mother’s training was reinforced at each follow up visit. The TrackCare app did not require any user input and would run in the background when the smartphone was turned on, or automatically restart if forced to stop by the user. Moreover, a password prevented its accidental or intentional uninstallation by any participant. The TrackCare app was configured to record and save the device location data (latitude, longitude and positional accuracy), source of the location data (GPS chipset, mobile network), location mode (high accuracy, device only, battery saving option enabled, location services disabled) and the timestamp, at one-minute intervals throughout the 6-month follow up period. The TrackCare app was configured to upload the saved location data and the upload timestamp at hourly intervals (or in case of poor network signal, at the next scheduled interval), to a secured central study server database which in turn would synchronize real-time with another secured mirror image server. The location data was cleaned to remove duplicate records and cached data (arising when the TrackCare app used the previous location data to infer the present location in the case of a weak GPS or network signal strength) and incorrect timestamps were adjusted, so restricting the data set to a single coordinate for each minute. When multiple geo-coordinates were recorded during the same time interval, the geo-coordinate from the most robust location data source was retained. GPS data with low accuracy due to poor signal strength or environment interference were removed. Missing values were interpolated between geo-coordinates that were less than an hour apart or within 100 m of each other [39]. The analysis for this paper was restricted to mothers from the phone group only, seeking care at qualified health care providers (practitioners of modern and Indian systems of medicine at hospitals, health centres and clinics) for their child’s illness. Care sought from Accredited Social Health Activists (ASHA), and Integrated Child Development Service (ICDS) workers was excluded as they typically provide peripatetic services once or twice a month and are accessed infrequently by mothers for curative care for child illness in the study setting. Pharmacies, drug-stores and shops were also excluded as infrequent secondary sources to access medicines following a visit to the health care provider. Furthermore, these sources often shared the same or adjoining location and are hence unsuited for separate detection from the health facility. The location data transmitted by the study phones was used to identify a temporal and spatial clustering of geo-coordinates around any of the health care facilities indicative of a care-seeking event based on four parameters – (1) phone proximity to health care facility (proximity range), (2) minimum time period spent in proximity of health care facility (minimum time), (3) maximum time period spent in proximity of health care facility (maximum time), and (4) time spent outside proximity of health care facility within this minimum and maximum period (time outside) to account for the random error in the location data (jitter) due to the varying signal strength and positional accuracy. Proximity range, minimum time, and time outside are analogous to parameters used to identify visited locations from participant trajectory data [44]. Maximum duration was included to eliminate likely non-visits arising from participant’s whose daily movement brings them within range of a health care facility for extended periods of time. Furthermore, health care facilities situated near the mother’s home were excluded for that mother, as it would not be possible to differentiate between the clustering of the location data around the mother’s house and the health care facility. In general, a care-seeking event was defined if the mother’s phone localized within a certain proximity of a health care facility (excluding those which are situated near ie, within a certain distance of the mother’s home) for a minimum and maximum period of time, further allowing for a small continuous period of time, for the phone to localize outside the proximity of the health care facility within this period. In the absence of a validated definition to identify a pattern of movement suspension or a trip to a health care facility based on GPS location data, we estimated the percent positive agreement for a care-seeking event (hereinafter referred to as concordance) ie, the agreement was the frequency with which the mother’s recall of a health facility visit matched with the location GPS data recorded by the TrackCare and defined as a potential visit using the parameters described above. We did not consider negative percent agreement (ie, agreement between the TrackCare and mother’s recall that a visit did not take place) in our concordance analysis as this would have falsely and highly inflated the overall concordance due to the large number of data points recorded by the TrackCare app outside the proximity of a health facility. Disagreement was also estimated between TrackCare and mother’s recall that a health facility visit did or did not take place. We did a sensitivity analysis using 6480 different threshold combinations for the various parameters to estimate concordance for the various parameters based on the GPS location data – proximity range (15 thresholds from 5 to 75 m with 5 m increments), minimum time (6 thresholds from 5 to 30 minutes with 5-minute increments), maximum time (6 thresholds of 2, 3, 6, 9, 12 and 24 hours), time outside (3 thresholds of 5, 10 and 15 minutes) and exclusion of health care facility if near mother’s home (4 thresholds of 50, 100, 200 m or no exclusion). We used multiple linear regression to model the effect of the parameters on concordance. The parameters with the different thresholds were treated as indicator variables to estimate the adjusted effect of each threshold of each parameter on the concordance between a mother’s recall and the TrackCare App for a care-seeking event. As a secondary objective, we also looked for concordance between TrackCare and mother’s recall of a care-seeking event specific to a calendar date. Within each level of the mother’s report, we analysed to see if the concordance varied by provider type (public or private), and type of health care facility (hospital or clinic). The study was approved by the Ethics Committee of the KEM Hospital Research Centre, Pune, India (Study No. 1415) and the University of Edinburgh, UK. Mothers provided written informed consent prior to enrolment and randomization. Prior to consent, mothers were informed that those assigned to the phone group would be allowed to keep the study phones even if they withdrew their participation at any stage of the study. Mothers in the phone group consented to the collection of their location data. The location data on the phone device was encrypted and erased as soon as it was transferred to the secured central study server. The study was not registered with the Clinical Trials Registry as it was not considered by investigators to meet the criteria for such a trial. Study groups differed in the method and frequency with which their care-seeking behavior was measured but no group was provided with a health-related intervention intended to affect a health outcome.

• hirve, siddhivinayak
• marsh, andrew
• lele, pallavi
• chavan, uddhavi
• bhattacharjee, tathagata
• nair, harish
• campbell, harry
• juvekar, sanjay