Comparison of next‐generation portable pollution monitors to measure exposure to PM2.5 from household air pollution in Puno, Peru

Abstract Assessment of personal exposure to PM2.5 is critical for understanding intervention effectiveness and exposure‐response relationships in household air pollution studies. In this pilot study, we compared PM2.5 concentrations obtained from two next‐generation personal exposure monitors (the Enhanced Children MicroPEM or ECM; and the Ultrasonic Personal Air Sampler or UPAS) to those obtained with a traditional Triplex Cyclone and SKC Air Pump (a gravimetric cyclone/pump sampler). We co‐located cyclone/pumps with an ECM and UPAS to obtain 24‐hour kitchen concentrations and personal exposure measurements. We measured Spearmen correlations and evaluated agreement using the Bland‐Altman method. We obtained 215 filters from 72 ECM and 71 UPAS co‐locations. Overall, the ECM and the UPAS had similar correlation (ECM ρ = 0.91 vs UPAS ρ = 0.88) and agreement (ECM mean difference of 121.7 µg/m3 vs UPAS mean difference of 93.9 µg/m3) with overlapping confidence intervals when compared against the cyclone/pump. When adjusted for the limit of detection, agreement between the devices and the cyclone/pump was also similar for all samples (ECM mean difference of 68.8 µg/m3 vs UPAS mean difference of 65.4 µg/m3) and personal exposure samples (ECM mean difference of −3.8 µg/m3 vs UPAS mean difference of −12.9 µg/m3). Both the ECM and UPAS produced comparable measurements when compared against a cyclone/pump setup.


| INTRODUC TI ON
Household air pollution (HAP) adversely affects nearly three billion people who use open fires and biomass fuels as their main sources of fuel for cooking. [1][2][3] Burning biomass fuels, particularly in poorly ventilated cooking areas, can emit dangerously high concentrations of pollutants such as particulate matter, black carbon, and carbon monoxide into the household environment. 2 A critical cut-point size of particulate matter is less than 2.5 micrometers in aerodynamic diameter (PM 2.5 ), due to its ability to penetrate lung tissue, traverse tissue barriers, and ultimately enter the bloodstream to cause systemic inflammation. 4 The World Health Organization (WHO) estimates that four million premature deaths and 45% of pneumonia deaths in children less than five years of age can be attributed to HAP exposure from biomass fuel use. 5 Pregnant women and their young children are especially vulnerable due to the large amount of time spent indoors cooking. 6 These include higher risk of low birthweight, increased stillbirth, 7,8 higher odds of acute lower respiratory infections among children less than five years of age, 9 and higher risk of perinatal mortality, neonatal mortality, and macerated stillbirths. 8,10 Few studies, however, have adequately measured personal exposure to household air pollution because of lack of easy-touse and easy-to-deploy equipment. 11,12 In past studies, intention-to-treat analyses lacking personal exposure measurements have led to exposure misclassification and underestimating the true effect of HAP exposure on health outcomes due to stove stacking. 13 Attempts to formulate exposure-response curves for HAP and personal exposure to PM 2.5 against respiratory health outcomes have been limited by the lack of data. Smith and Peel 14 drew attention to the lack of HAP personal exposure data used in exposure-response analyses. For example, the integrated exposure-response analyses developed by Burnett et al 15 uses only a limited amount of HAP data. As of today, the situation has not greatly improved. Overall, systematic measurement of personal exposure to PM 2.5 in conjunction with specific health outcomes is strongly needed to increase precision of exposure-response estimates and better evaluate the impacts of cookstove and clean fuel interventions in the future.
There continues to be poor assessment of personal exposure in previous cookstove studies. For example, the Randomized Exposure Study of Indoors and Respiratory Effects Study (RESPIRE) in Guatemala did not measure personal PM 2.5 exposure measurements and used carbon monoxide (CO) exposure instead to approximate particulate matter exposure rather than direct PM 2.5 exposure assessment. While the relationship between CO and PM 2.5 was well-validated for the specific setting of the Western Highlands in Guatemala, this relationship has been shown to have poor-to-moderate correlation in other contexts. 16,17 Traditional pump and cyclone approaches for assessing personal exposure include a battery-powered sampling pump typically worn on the hip or in a backpack with tubing attached to a particle size-selective device installed in the human breathing zone. 11,18 These setups are generally expensive, noisy, heavy, and bulky for personal exposure measurements on women and children 19,20 and are often filter-based, requiring laboratory analysis and delaying results for study participants and subsequent analyses. Therefore, there is a strong need to validate lighter, lower-cost, and easier-to-use tools for air quality monitoring in low-resource field settings. 21,22 These instruments will thus improve measurement of an individual's personal air pollution intervals when compared against the cyclone/pump. When adjusted for the limit of detection, agreement between the devices and the cyclone/pump was also similar for all samples (ECM mean difference of 68.8 µg/m 3 vs UPAS mean difference of 65.4 µg/m 3 ) and personal exposure samples (ECM mean difference of −3.8 µg/m 3 vs UPAS mean difference of −12.9 µg/m 3 ). Both the ECM and UPAS produced comparable measurements when compared against a cyclone/pump setup.

K E Y W O R D S
exposure assessment, fine particulate matter, household air pollution, instrument validation, lower-and middle-income countries, personal exposure

Practical Implications
• We conducted this study in Puno, Peru, to evaluate the ability of two exposure assessment instruments to collect particulate matter concentrations from household air pollution and compare them against a traditional cyclone and pump method.
• These relatively less expensive, lighter, and easier-to-use instruments, as compared to traditional instruments, may enable higher spatiotemporal resolution for air quality monitoring in real-world field settings.
• Thus, this will help improve exposure-response estimates for health outcomes and exposure classification, and inform health program planning, future intervention evaluation, and chronic disease management. exposure and exposure classification in their home, exposure-response estimates for health outcomes, and ultimately data resolution in low-to middle-income countries for health program planning, future intervention evaluation, and chronic disease management. 23

| Study setting
The Household Air Pollution Intervention Network (HAPIN) trial, the parent study to the work described here, is a randomized controlled trial of a liquefied petroleum gas (LPG) stove and continuous fuel distribution in 3200 households in four LMICs (India, Guatemala, Rwanda, and Peru) (https ://clini caltr ials.gov/ct2/ show/NCT02 944682). This pilot study took place in Puno, Peru, which lies at an altitude of 3825 meters above sea level and has low temperatures (average range of 5.9°C-9.8°C) throughout most of the year. 24 Most communities in this region speak Spanish, Aymara, and Quechua languages, and use wood, dung, and crop residues as cooking and heating fuels in traditional stoves (known as fogón in Spanish).

| Study design
We sought to test the wearability of two new devices and agreement with a cyclone/pump PM 2.5 exposure assessment equipment.
The Enhanced Children MicroPEM (ECM, RTI International) is a recently developed instrument that is a relatively lower-cost, easier-to-use PM 2.5 gravimetric, and real-time exposure assessment tool when compared to a cyclone/pump setup ( Figure 1A). The ECM is the size of a pack of cards, weighs about 150 grams (RTI International), has a flow rate of 0.3 L/min, short battery charging times (4-6 hours), and uses small 15 mm filters. This instrument can also measure particle concentration via nephelometric particle scattering, participant compliance with an accelerometer, and it logs temperature, relative humidity, and flow rate. An accelerometer is useful in measuring participant compliance in wearing sampling instrumentation, as it can detect participant activity levels to confirm that the participant is physically wearing the device. 25,26 The second instrument evaluated in this study, the Ultrasonic Personal Air Sampler (UPAS; Access Sensor Technologies, Fort Collins, CO) ( Figure 1B), weighs 230 grams, has a flow rate of 1 L/ min, can be charged within three and a half to four hours, and is about the size of a standard smartphone (128 × 70 × 33 mm). The UPAS collects PM 2.5 data gravimetrically on 37 mm filters and collects motion data via an accelerometer, temperature, and humidity data. It can also be programmed to start sampling within certain GPS coordinates and has an Android/iPhone application for instrument configuration, flow rate, programming of sample characteristics, and downloading of data (UPAS v2.1.9; Access Sensor Technologies). Both instruments were used to collect filter-based PM 2.5 samples to compare against gravimetric samples collected by a currently accepted traditional pump and cyclone setup. In this study, we used a Triplex Personal Sampling Cyclone (Mesa Labs) with a SKC AirChek XR5000 pump (SKC) with a weight of 450 grams and sampling at 1.5 L/min. This setup has been used frequently in previous validation studies 18,25,27,28 (Figure 1C).

| Exposure assessment
Eighty-two co-location PM 2.5 HAP measurements were collected with the three co-located instruments (ECM, UPAS, and the cyclone/ pump). To evaluate the range of PM concentrations typical of rural households using biomass or gas stoves, we used three different classifications of 24-hour HAP exposures (low, medium, and high pollution environments, as described below) and collected PM 2.5 gravimetrically with all three instruments. 29 The primary motivation for this study was to compare the agreement of the ECM and UPAS with the cyclone/pump setup at low-exposure concentrations that can contribute to the lower end of the exposure-response curve, as little data currently exists for this range. 11,14,30 Previous studies looking at different airborne particulate exposures and similar health endpoints have conflicting findings, leading to a recommendation to improve exposure classification to define specific exposure cut-points on the response curves that may help design better interventions and formulate policy. 11,14,30 Additionally, these lower exposures may be representative of 24hour exposures in homes using clean fuels. To the best of our knowledge, this will be the first time this range of concentrations has been evaluated using these instruments.
To compare PM 2.5 measurements for low exposures, or concentrations that may be representative of exposures in homes using intervention gas stoves, 39 field workers had three instruments (ECM, UPAS, and cyclone/pump) co-located in the middle of their breathing zone (defined as on the chest halfway between the throat and diaphragm) while wearing a custom-made kitchen apron with special pockets to hold the instruments together and keep the air inlets unobstructed and exposed ( Figure 1D) to air.
Field workers turned on instruments prior to entering study households and continued to have the machines run throughout study visits. Throughout the sampling period, field workers recorded the number of study households visited during daily follow-ups, time spent in each household, and type of cookstove in the households.
Low exposures were defined in our study as a mixture of short 5-10-minute visits spent in gas and/or traditional cookstove homes.
Upon completion of study visits for the day, the field workers returned to the office and left the study apron hanging on coat hooks in the office overnight with all instrumentation inlets exposed until the machines turned off after 24 hours of sampling. All instruments (ECM, UPAS, and cyclone/pump) were run at 100% duty cycle, or continuous sampling.
To evaluate performance at medium levels of exposure (ie, HAP exposure levels representative of a typical biomass-using primary cook in a Peruvian household), five study participants with daily biomass fuel use were enrolled and wore the same personalized study apron with the co-located instruments that field workers used for low-exposure assessment ( Figure 1E). A total of 15 co-location samples were collected. Consent forms were collected and signed at the time of enrollment, and participants were instructed to wear the study aprons for a 24-hour time period, and to take off and place the aprons in the same room and breathing zone as themselves whether they were bathing or sleeping. 29 The duty cycles on the ECM's were changed to 30 seconds on, 30 seconds off (50% duty cycle) to avoid inlet blockages and battery life failures based on previous ECM testing results (unpublished data). The duty cycle on the UPAS and pump and cyclone remained at 100% (continuous sampling) due to the larger filter size used in both samplers (37 mm).
To evaluate performance at high levels of exposure (ie, exposures representative of large HAP concentrations from staying in a kitchen for an entire cooking event), birdcages containing an ECM, UPAS, and cyclone/pump were installed in a kitchen at one meter horizontally and 1.5 m vertically in height from the cookstove combustion zone. 29 A total of 28 co-location samples were collected. An example installation is shown in Figure 1F. We visited each of these households and collected their area concentration measurements three times over the study period for repeated measurements, since we assumed that day-to-day sampling variation was sufficient for independent samples. 20 Due to the high pollution concentrations in the kitchens, the ECM duty cycle was changed to 20 seconds on, 160 seconds off (11.1% duty cycle) to avoid inlet blockages and battery life failures while the UPAS and cyclone/pump remained at 100% duty cycle as based on previous ECM testing results (unpublished data).

| Sample and filter quality control
Flow rates for all instruments were measured and recorded before and after collecting each sample in the field. Field blanks, that is, filters placed into a sampler and installed in the household for 24 hours, while the sampler is off, were collected at a frequency of five times for every two weeks of sampling to account for background contamination that might occur during filter installation. Duplicate samples, that is, co-located filters in two of the same type of instrument, were collected once out of every 10 area samples collected. Duplicates and field blanks were not able to be collected on study participants, as the additional weight of the apron would have likely affected participant compliance. Lastly, laboratory blanks, that is, filters that remain in the laboratory and are not deployed to the field, were also collected and sent with each batch of filters for analysis.

| Filter assessment
Prior to being sent to Peru, filters were first conditioned in a climate-controlled environment at the University of Georgia, Athens (UGA) (temperature range: 19.9-20.1°C; relative humidity range: 39.7%-42.5%) per Environmental Protection Agency laboratory guidelines. 31,32 In brief, the filters were placed in individual Petri dishes, laid out flat in the analysis clean room, and lids were cracked slightly ajar to allow slight airflow but not enough to allow particle/dust contamination to the filters. After conditioning in these settings for 48 hours, the filters were pre-weighed prior to transport to Peru. After sampling in Puno, filters were stored in a freezer and then sent back to UGA for analysis using ice packs to keep the filters cool during transport. The filters were respectively. 33 The gravimetric mass weights were then field blank weight-corrected by subtracting the median field blank filter weight, as well as LOD-corrected using the LOD mass weights described above.
In brief, any filter mass measurements not exceeding the limit of detection were replaced by LOD∕ √ 2 in the calculation of the gravimetric mass concentration. 34 If blank-corrected, gravimetric filter weights fell below the LOD, they were excluded from the main analysis, but all LOD-corrected values were used in the sensitivity analysis.

| Biostatistical methods
The primary analytical objective was to assess agreement between PM 2.5 concentrations measured by the ECM and the UPAS compared to the cyclone/pump method. This primary comparison analysis was done by using the Bland-Altman method 35 to assess agreement and estimate bias in measurements. The Spearman correlation was calculated for both instruments to assess correlation 36,37 due to the non-parametric nature of the collected data. These statistical analyses were first done using all data points, and then, further stratified by sampling environment, that is, high kitchen HAP concentrations versus low and medium personal HAP exposure measurements, to assess agreement and correlation. We performed these analyses with the assumption that all collocated instrument assessments were independent of each other and that the day-to-day sampling variability in PM 2.5 measurements was enough to support this assumption. 20 We conducted a sensitivity analysis where we included only co-locations in which all three sampling instruments had no runtime or filter issues and in which all three instruments simultaneously collected samples with filters above the LOD to see how this selection would affect results. The statistical programs STATA Version 15 (StataCorp) and R (R Foundation, www-rproj ect.org) were used to conduct these analyses.
Finally, as mentioned above, one advantage of the ECM device over the UPAS is its ability to collect and display real-time nephelometric data that can be then be corrected against gravimetric filter data. 27 Using the real-time data collected by the ECMs, we were able to plot the hourly distribution of concentrations over the 24-hour day collection period during this pilot study against the gravimetrically corrected real-time data in each of the three exposure settings. 38 Table S1, and the reasons for filter flagging and dropped samples are included in Table S2.

| Sample characteristics
There were no differences in sampling times for all samples (P = .72), area samples (P = 1.00), or personal samples (P = .66).

| Average concentrations
All descriptive statistics by instrument are described in Table S3. In brief, all three instruments collected similar concentration ranges and averages during the 24-hour co-location experiments. All three instruments were able to collect samples over a large range of concentrations but also could handle very high particulate concentrations that are typically observed in indoor biomass-user kitchens in our study area, as well as detect low concentrations. For the field blanks collected for each instrument, the median mass loading of these blanks was 1.5 µg, −3.2 µg, and −4 µg for the ECM, UPAS, and cyclone/pump, respectively.

| Agreement between devices
We present correlations and agreement between each next-generation device and the cyclone/pump in   In terms of bias, the UPAS results did not appear to differ significantly between the main and this second sensitivity analysis but did However, even with these slight changes in Spearman's correlation values and magnitude of biases, all the 95% confidence intervals overlapped between the main and sensitivity analyses and the mean differences were not significantly different from one another (data not shown), 39 indicating that there were no significant differences observed between the two analyses. We summarized these findings in Table S6.

| Real-time assessment
Using real-time data collected by the ECMs, we plotted an hour-byhour summary over the 24-hour collection period against the gravimetrically corrected real-time data in each of the three exposure settings. As an advantage of using the ECM over the UPAS, as displayed in Figure S1, we were able to visualize and summarize temporal cooking patterns of our participants, and peak exposure times and concentrations throughout the day. This figure shows peak concentrations of exceeding 500 µg/m 3 during cooking hours, at around 6 am and 6 pm.

| D ISCUSS I ON
We collected 24-hour gravimetric samples of an ECM and UPAS co-located with a traditional cyclone/pump to measure air concentrations and personal exposures. Overall, we found that in our study setting of Peru, both the ECM and UPAS performed well and measured comparable concentrations to those from the traditional gravimetric instruments. The overall bias in either device undergoing testing was estimated to be approximately less than 10% of the measurement concentration range obtained from the traditional gravimetric device (ECM overall bias: 121.7 µg/m 3 , UPAS overall bias: 93.9 µg/m 3 out of 1303.2 µg/m 3 range of concentration measurements for the pump and cyclone). This bias is within range for these types of air sampling instruments, as calibration comparisons of the traditional pump and cyclone used in this study compared to reference standards sampling at similar flow rates and cut-points have been shown to have a bias of 10% between themselves. Within the two sets of duplicate cyclone/pump samples collected for this study, the percent differences in concentrations collected were approximately two percent and seven percent (data not shown).
In this study, we performed a co-location comparison of the ECM and UPAS against a traditional gravimetric instrument commonly used in field studies. However, past studies have sought to evaluate both laboratory and field performance of just one of these instruments or past versions of the instruments compared to commonly used gravimetric instruments. 18,25,[40][41][42][43][44][45][46][47][48] Although there are no evaluation studies to date that have been conducted on the current version of the ECM, except for a pilot study to compare the ECM to the MicroPEM, 40  The potential explanations for these differences may be due to the same reasons described for the ECM measurements above, but future evaluation studies for the UPAS are also needed.
Our study had several strengths. We evaluated two new air quality instruments, the ECM and the UPAS, against a traditional gravimetric field instrument over a large range of personal exposure and area concentrations. As cited above, past studies have typically only evaluated one of these instruments against gravimetric standards and in laboratory settings. By conducting this study in the field, this gives us a better understanding of real-life performance 21,22 in low-resource settings with non-technical field staff, potential is- types, any single monitor may have been influenced by placement in the kitchen (ie, one inlet is more exposed than others, even though they are close together), as well as instrument operation-specific issues. Third, personal exposures (both "low" and "medium" as designated in our manuscript) are typically lower, sometimes much lower, than area concentrations, thus leading to a greater number of samples being represented in these lower concentration ranges compared to higher area concentrations. Finally, this analysis was only conducted using data from Peru and thus more testing needs to be conducted in other LMIC's.
In terms of the differences in results observed between the main and sensitivity analyses, potential explanations for these changes in correlation and bias are four-fold. During this period of testing, based on what we had observed for biomass stove area monitoring and ECM filter clogging issues in our previous CHAP study, 29 we instituted 50% (primary cook monitor) and 11.11% (area monitor) duty cycling in the ECM's that were installed in biomass-using homes.
However, because this study only conducted 24-hour sampling compared to CHAP's 48-hour sampling regimen, this shortened amount of sampling time may explain why so many of the ECM's filters fell below the LOD cutoff (n = 33, 45%) compared to the UPAS (n = 22, 31%) and the cyclone/pump (n = 7, 10%), as well as the variability in the household sampling time necessary to obtain the LOD saturation exposure. Second, due to the ECM's filter size (15 mm) in combination with the duty cycling, compared to the larger UPAS filter (37 mm) with 100% duty cycling for all samples, the differential testing regimen between the two instruments for only 24 hours of sampling may have led to insufficient mass loading for the ECM.
Third, it is unknown whether the 15 mm filter size may have clogged sooner than the 37 mm filters used in the cyclone/pump setup and the UPAS due to smaller surface area in highly polluted kitchen environments. However, since the amount of bias was similar between the ECM and the UPAS, the filter size may not have been a major issue overall. Lastly, in our particular field setting, we saw that the staff had had more experience using the ECM's compared to the UPAS due to their familiarity with using them in our CHAP study. 29 Due to the novelty and unfamiliarity of using the UPAS in our field setting, more attention may have been paid to the filter handling and processing steps of cleaning the UPAS compared to the ECM, as observed by the number of UPAS samples above LOD.

| CON CLUS IONS
The ECM and the UPAS both performed well overall and produced comparable agreement and mean difference statistics with overlapping confidence intervals to the same traditional gravimetric sampling instrument. Due to the sample size collected in our study, more data would be helpful to determine the performance of these instruments at lower concentrations.

ACK N OWLED G EM ENTS
The HAPIN trial and this pilot study are funded by the US National

N E T WO R K (H A PI N) I N V E S TI G ATO R S (LI S TED I N A L-PH A B E TI C A L O R D ER )
• HAPIN Steering Committee: Kalpana Balakrishnan (Sri