Decision tools for schools using continuous indoor air quality monitors: a case study of CO₂ in Boston Public Schools

Summary

Background

Indoor environments in schools affect children’s health and learning, but without indoor air quality (IAQ) monitoring, staff have limited data to inform interventions. We aimed to demonstrate the utility of comprehensive real-time classroom IAQ monitoring for informing interventions and investments in schools.

Methods

We created analytical datasets of classroom carbon dioxide (CO₂) from a groundbreaking network of 3659 commercial IAQ monitors installed across a U.S. school district. We analysed patterns and trends in 125 school buildings over the 2022–2023 school year. Working with Boston Public Schools (BPS) staff, we co-created visualisations suitable for school decision-making and for research.

Findings

Over the school year, 245 million CO₂ measurements were recorded. The average school-day CO₂ concentration was 841 ppm (SD: 405 ppm) and the 99.9th percentile of daily maxima was 5080 ppm. Concentrations varied by over an order of magnitude within and between schools. Elevated levels occurred in all buildings, although schools with central mechanical ventilation had lower concentrations and variability than schools without. We show multiple visualisations of monitoring data using threshold-based exposure metrics.

Interpretation

Our academic-school partnership enabled actionable research in support of IAQ improvements. The heterogeneity of CO₂ levels across BPS classrooms and between schools underscores the need for continuous classroom-level monitoring and demand-controlled ventilation. Results and protocols are generalisable to other schools and researchers. Our work paves the road for future studies of IAQ and student health and learning.

Funding

Boston University Initiative on Cities, Boston University School of Public Health, National Institutes of Health, National Science Foundation.

Research in context.

Evidence before this study

We searched PubMed for studies published before November 1, 2024 reporting carbon dioxide (CO₂) concentrations in primary and secondary schools worldwide using the following query: “school∗” AND “classroom∗” AND (“carbon dioxide” OR “CO2” OR “ventilation”). We found few examples of comprehensive, year-long indoor air quality (IAQ) monitoring campaigns conducted in schools. Most studies reporting on schools’ IAQ were conducted over short periods of time (hours ∼ months). Of the limited number of studies reporting the results of continuous monitoring for at least the length of a school year, none were conducted for all classrooms across entire school buildings or across all schools in a district. Moreover, only a subset of studies provided detailed descriptions of procedures for assessing data quality and validating sensor measurements. Finally, we identified minimal evidence of meaningful school engagement in guiding research enquiries or using monitoring data to inform interventions. This aligns with findings that misaligned researcher and school goals, and funding entities’ lack of interest in applied research can hinder effective academic-school partnerships.

Added value of this study

The academic-school partnership described in this study provides a model for conducting actionable research that serves common interests. To our knowledge, this study is the first to report data from a year-long CO₂ monitoring campaign spanning an entire school district. We leveraged a groundbreaking network of 3659 sensors to characterise fine-scale variations in CO₂ concentrations across each school day in the 2022–2023 school year for each classroom in Boston Public Schools (BPS). We found evidence of substantial variability in concentrations during hours of school operation, and even greater variability comparing between classrooms and schools across the school year. Mechanical ventilation status was a modifier of overall trends in concentrations, but there were still recorded instances of high CO₂ levels in classrooms in mechanically ventilated schools. We demonstrate the utility of continuous monitoring for assessing compliance with health- and comfort-based guidelines and informing subsequent interventions. We also present protocols for evaluating and improving data quality, which are critical when using data from commercial-grade sensors for both scientific research and school decision-making. Our work directly supports school district interests, given that school staff were involved with every step of this project.

Implications of all the available evidence

CO₂ concentrations recorded in BPS schools fell within the range of concentrations that has been reported previously. The levels of temporal and spatial fluctuations in the BPS data suggest that there may be misclassification if spot check data are used for exposure characterisation or epidemiological studies (especially those where measurements are intended to represent chronic exposures). Therefore, we advocate for the use of data from long-term monitoring campaigns in future studies examining the association between schools’ IAQ and students’ health and learning, as well as to support the development of evidence-based IAQ standards. However, such data must be subject to rigorous quality control given the potential for high variability, missingness, residual calibration bias, and sensor drift when utilising commercial-grade sensors. Our findings also support the role of mechanical ventilation in maintaining good IAQ and we discuss the added benefit of demand-controlled ventilation. Overall, our work highlights the importance of large-scale IAQ monitoring and sustained partnerships between academics and school staff. By expanding schools’ analytic capacity, academics can facilitate applied research in support of a healthier environment for occupants.

Introduction

Over 50 million students in the United States spend more time in school buildings than anywhere else besides at home,^1,2 where they are exposed to indoor environmental characteristics that can affect their health and learning.³ In the absence of sufficient regulation and funding, school facilities are often inadequately maintained,⁴ particularly in high-poverty school districts.⁵ Structural issues such as leaking roofs and malfunctioning heating, ventilation, and air conditioning (HVAC) systems can contribute to poor indoor air quality (IAQ).^4,6 Schools must optimise investments and operations to improve IAQ, yet stakeholders are rarely equipped with information to do so.⁶

The COVID-19 pandemic prompted a surge of interest in IAQ given its importance in driving infection rates,⁷ and schools pursued a range of IAQ-related activities to facilitate safe in-person learning after an initial wave of closures. Many implemented ventilation or air cleaning improvements, e.g. by using higher efficiency filters in HVAC systems and distributing air purifiers.7, 8, 9 Some installed sensors to measure key attributes of IAQ in real-time.^8,10 Monitoring empowers schools to take action when IAQ is poor and assess the efficacy of these interventions, yet few school districts in the U.S. conduct consistent and comprehensive IAQ monitoring.¹¹ Moreover, school staff are often not well-equipped to manage, analyse, interpret, or use monitoring data.⁸

Carbon dioxide (CO₂) is an IAQ parameter that reflects the balance between occupants’ respiration, ventilation, and outdoor CO₂.¹² Since the COVID-19 pandemic, organisations worldwide have issued CO₂ and/or ventilation guidelines to mitigate risk.¹³ Ceteris paribus, higher levels of CO₂ are associated with lower air exchange rates (AERs) and thus poorer air quality, increasing the risk of airborne infectious disease transmission as well as occupant discomfort (e.g., via perceptions of body odour).^14,15 The direct health effects of CO₂ at high concentrations are well-established, with a U.S. workplace workday time-weighted average limit of 5000 ppm.¹⁶ Some findings suggest that CO₂ may also be an indoor air pollutant at moderate concentrations, having been linked to impaired decision-making in some controlled experimental studies at concentrations as low as 1000 ppm^13,16 – levels that have been regularly recorded in classrooms.¹² Although more evidence is needed to confirm the impact of CO₂ exposure at moderate levels,^17,18 particularly in school settings, improved ventilation has been associated with improvements in student health, attendance, and performance.¹² Ventilation – whether natural or mechanical – dilutes indoor air pollutants and reduces exposure.¹² Since the gold standard (i.e., the tracer gas method) for measuring ventilation rates (VRs) is costly and impractical in classroom settings,¹⁹ measuring and controlling CO₂ concentrations as an indicator of VRs has been a priority for schools. However, the paucity of evidence-based standards for indoor CO₂ levels in the U.S. and the improper use of data for enforcing existing standards pose challenges for effective interventions.²⁰

Boston Public Schools (BPS) is the first school district worldwide to install and share IAQ data from thousands of IAQ monitors.¹¹ These data are used to support school decisions and conduct research in partnership with Boston University (BU). In this paper we report the results of measuring CO₂ in BPS schools. Since this is, to our knowledge, the first study of its kind, we present here 1) how the partnership was established, 2) protocols for transforming commercial CO₂ sensor measurements into analytical datasets, 3) a comprehensive overview of CO₂ levels and trends in BPS schools, and 4) how these datasets can be used as decision tools to improve IAQ.

Methods

Location

Boston Public Schools is an urban school district in Boston, Massachusetts serving around 46,000 pre-kindergarten to twelfth grade students (typically ages 3–18) across 119 schools.²¹ BPS is the oldest²² and one of the most diverse public school systems in the United States.²¹ In the 2022–2023 school year, 70% of BPS students were low-income²¹ and its sociodemographic diversity was representative of the nearly 15 million students attending urban public schools in the U.S.¹ The average class size was 16 students.²³ Over half of BPS buildings were built before the 1940s.²² Less than a third have central HVAC systems²²; the rest use steam heat and largely rely on operable windows and manually operated window air conditioning units, although a small subset has no mechanical cooling.¹⁰ Per Massachusetts workplace regulations, heating must be available in school buildings from October 15th to May 15th each year (n = 127 school days). In schools with central HVAC, the cooling season is not strictly defined, depending more on outside temperatures and building needs. Here, we designated the cooling season as lasting from September 8th–30th and May 16th-June 23rd (n = 44 days). There is a brief shoulder season between the cooling and heating season (n = 9 days). BPS schools vary in the extent to which temperatures in classrooms are centrally, zonally, or individually controlled.

Schools with central HVAC are mechanically ventilated when the HVAC system is operating to maintain temperature setpoints during the heating and cooling seasons, with adjustable proportions of fresh air vs. recirculated air. Only three school buildings (two with demand-controlled ventilation [DCV], in which mechanical ventilation is linked to CO₂ thresholds, and one without DCV but with central HVAC) of 125 had classroom-level HVAC ventilation controls as of the 2022–2023 school year. Some schools without central HVAC have limited supplemental mechanical ventilation (exhaust fans or unit ventilators with placements in some or all classrooms, but with windows remaining the primary source of ventilation) while others depend entirely on natural ventilation.¹⁰

Project history and partnership formation

Before the COVID-19 pandemic, BPS performed periodic spot-checks of IAQ parameters in accordance with its EPA Tools for Schools-based IAQ Management Plan through its annual School Environmental Audits programme. Interest from Facilities staff and mounting pressures regarding safe re-openings led BPS Facilities Management to successfully advocate for the installation of continuous monitors, funded with $6.7 million in Elementary and Secondary Schools Emergency Relief (ESSER) funds from the City of Boston. Between September 2021 and April 2022, BPS installed over 4000 sensors in classrooms, main offices and nurses’ offices, and on building rooftops that measure and record six indoor environmental quality parameters (CO₂, carbon monoxide, temperature, relative humidity, PM₁₀, and PM_2.5) every minute.¹⁰ In January 2022, BPS launched a public-facing dashboard featuring real-time 15-min rolling averages.¹⁰ BPS has used indoor CO₂ measurements as a marker of ventilation and to evaluate compliance with guidelines from its IAQ Monitoring and Response Action Plan.²⁴

The BU and BPS teams identified a shared interest in leveraging the sensor data for school decision-making, while the BU team recognised the potential for research supporting school operations and policies, and public health research. BU followed standard BPS protocols to formalise the research partnership. Since the approval of the study proposal, the BU research team has collaborated with BPS Sustainability, Energy, and Environment Programme staff and employees of SGS Galson (East Syracuse, New York), the sensor vendor. The entire team meets virtually on a biweekly basis.

Indoor CO₂ measurements and data storage

CO₂ levels were recorded on SmartSense monitors (SGS Galson, East Syracuse, NY), which measure CO₂ via Alphasense infrared sensors with an accuracy of ±10% of reading. We analysed the subset of sensors (n = 3659) installed by vendor technicians across all BPS classrooms (one per classroom), placed on internal walls at a height of 4–8 feet, away from windows and vent outlets where possible. The vendor maintained the monitors year-round and performed calibration annually in the field during summer breaks and the start of the school year using a two-point calibration method, with a 1000 ppm CO₂ reference gas, and a comparison ambient measurement recorded with a calibrated Q-Trak Indoor Air Quality Monitor (TSI, Shoreview, MN). Minute-level data were recorded and transferred to BU through an encrypted Amazon Web Services Simple Storage Service bucket. Data were stored and analysed in a PostgreSQL database on the BU Shared Computing Cluster, a high-performance heterogeneous Linux cluster.

CO₂ data were linked to information about buildings’ ventilation systems (classified as central, supplemental, and none), school-specific opening and closing times, and district-wide no-school days obtained from BPS. No-school days across the district included federal holidays and additional recesses, mainly winter recess (December 23rd-January 3rd), February recess (February 20th–24th), spring recess (April 17th-21st), and summer recess (prior year’s ending September 7th, 2022 and study year’s beginning June 24th, 2023). Unless specified otherwise, data presented here were restricted to school days and operating school hours. Classrooms of all grade levels within the same school building had the same start and end times. The average school-day length was 6.7 h (SD: 0.6 h) and class blocks were generally 40–60 min long with short breaks (∼5 min) in-between in addition to one lunch break, although the exact timing of breaks was not available. In this analysis, we present data collected from September 8th, 2022 to June 23rd, 2023.

Data completeness, cleaning, validation, and sensor sample size

We estimated overall data completeness by comparing the number of actual observations (excluding duplicates and missing values, and standardising to minute-level) and expected observations (total minutes per school year). We evaluated trends in missingness over time.

We generated an offset correction for each sensor’s school-day CO₂ measurements to reduce residual calibration error. For each school, we calculated a monthly baseline CO₂ value by averaging nighttime (12 am–4 am) CO₂ measurements collected across all sensors, excluding outliers outside the 1st–99th percentiles. We then calculated the same monthly nighttime average per sensor and used the difference between monthly school and sensor nighttime averages as an offset correction for each respective sensor.

We calculated a 15-min rolling median to smooth fluctuations and impute missing values. Lastly, we excluded values outside 370–8700 ppm based on data percentiles, accepting that some uncontrolled residual error remained, to produce our final adjusted CO₂ dataset. Supplemental Fig. S1 summarises the cleaning steps described above.

We assessed the impact of offset adjustments by summarising and plotting the distributions of nighttime measurements before and after adjusting, and by comparing the distributions to typical outdoor levels. We defined these levels as 370–500 ppm based on a prior monitoring campaign conducted with a rooftop instrument tower located in a heavily developed urban area in Boston.²⁵ To evaluate sensor stability and identify potential drift, we modelled linear trends in nighttime average concentrations across the school year (including weekends and holidays) assuming daily nighttime equilibration with outdoor levels. We validated the sensors by comparing their CO₂ measurements in nine unoccupied classrooms spiked with dry ice with those of a calibrated Q-Trak Indoor Air Quality Monitor (TSI, Shoreview, MN). We calculated the correlation (r) and absolute difference between the minute-level readings.

We also leveraged the CO₂ data to demonstrate the extent to which differently sized samples of sensors would adequately characterise a building’s mean CO₂ levels. For each sample size and each school building, we drew 100 random samples and calculated the absolute percent difference between the sample school-year mean concentration and the census mean. Differences of greater than 10% reflect biased sample means given the sensor specifications. We further stratified our results by size of school building to provide recommendations.

CO₂ concentrations and metrics

We computed overall summary statistics for CO₂ concentrations (including coefficients of variation [CV]) across and within BPS schools; by ventilation status; and by heating, cooling, and shoulder seasons. We also calculated the percentage of the variability in one classroom’s concentrations that could be explained by another classroom’s concentrations (r²) for classrooms from the same school building, and classrooms’ pairwise correlations. We calculated daily maximum and mean concentrations, and the percentage of each school day where concentrations exceed certain thresholds as described below. For daily metrics, we only analysed days of data with at least 95% completeness.

Data visualisations and applications

We worked with BPS staff to generate visualisations and analytical products in support of funding applications, presentations to school administrators, operations and maintenance, and Facilities staff's decision-making. We compared calculated metrics with threshold values from the BPS IAQ Monitoring and Response Action Plan, which categorises maximum indoor CO₂ concentrations of 400–1000 ppm as “typical”, 1000–2000 ppm as “moderate”, and above 2000 ppm as “high”.²⁴ These thresholds are framed as comfort guidelines rather than public health standards, although the Plan states a goal of maintaining CO₂ within the “typical” range for “improved indoor air quality and a healthier school environment”.²⁴ We also demonstrate for illustrative purposes the use of evidence-based thresholds drawn from a recent review of worldwide CO₂ guidelines.²⁰ We use a 941 ppm ceiling limit for pandemic periods, as derived by the Nordic Ventilation Group (NVG) in 2023 to maintain an acceptable risk of infection in classrooms with a floor area of 42.5 m², a room height of 2.9 m, 25 occupants, and an occupancy duration of 6 h. For occupant comfort during non-pandemic periods, we applied the multi-tier limits proposed by the European Committee for Standardisation (CEN) in 2019, corresponding to VRs needed to control odours from human bioeffluents to specified levels of occupant satisfaction (950 ppm: 15% dissatisfaction; 1200 ppm: 20% dissatisfaction; 1750 ppm: 30–40% dissatisfaction).

All analyses and visualisations were done using R (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria).

Ethics approval

This study was not human subjects’ research, and therefore Institutional Review Board approval was not required.

Role of the funding source

The study sponsors had no role in the study design; data collection, analysis, or interpretation; writing the report; or deciding to submit the paper for publication.

Results

Data completeness

During the school year, 3659 classroom sensors collected over 245 million minute-level CO₂ measurements across 125 BPS school buildings. On average, sensors recorded at least one observation during school hours on 95% of days (171 days) throughout the school year (SD: 14% or 25 days). Sensors recorded an average of 92% (SD: 16%; range: 4%-99.99%) of expected observations. 3.5% of entries (9 million) were removed due to being duplicates or having missing CO₂ measurements. Smoothing reduced missingness to 5.3%. Overall, there was no systematic change in missingness over the school year.

On average, 0.3% (SD: 1.2%, range: 0–23%) of each sensor’s 15-min rolling median daytime CO₂ measurements were excluded due to falling outside the filtering range. Restricting to complete days of data resulted in an average retention of 98% of days per sensor (SD: 6%; range: 5%–100%). Supplemental Fig. S2 demonstrates how data completeness varies across a sample of 20 random schools over the school year.

Offset adjustments & drift

Offset adjustments using monthly nighttime averages ranged from −262 to 418 ppm. Supplemental Fig. S3 shows distributions of raw and adjusted nighttime CO₂ concentrations from a random sample of sensors in one school. The adjusted nighttime concentrations are more centred around the school mean nighttime concentration and more normally distributed than their raw counterparts and fall within the expected range given typical outdoor levels. The average slope from regressing daily sensor nighttime CO₂ averages on time was 0.10 (SD: 0.84), indicating an upward trend. Around half the sensors had slopes corresponding to drift exceeding the sensors’ margin of error assuming equilibrated nighttime concentrations; a quarter had drifts exceeding 100 ppm.

Data validation

We found high correlations in minute-level measurements (0.95–1) between the SmartSense and Q-Trak sensors, and an average absolute difference of 72 ppm (SD: 74 ppm). There was no systematic pattern of the SmartSense measurements being higher or lower than the Q-Trak measurements.

The impact of sensor sample size

One percent or less of sample means was unbiased with the use of 15 sensors for buildings with 16–30 classrooms, 20 sensors for buildings with 31–45 classrooms, and 25 sensors for buildings with 46–75 classrooms. For all building sizes, using five sensors yielded biased mean estimates across at least ten percent of samples. Supplemental Table S1 summarises the percentage of biased sample means by sample size and stratified by the number of classrooms per school building, and Supplemental Fig. S4 shows the absolute percentage differences between the sample and census means.

Nighttime summary statistics

The average nighttime CO₂ concentration on school days in SY22-23 was 460 ppm prior to applying offset corrections, and 485 ppm after adjustments. Overall, 80% of raw nighttime measurements and 92% of adjusted measurements fell within typical outdoor levels.²⁵ The average raw nighttime measurement for 86% of sensors was within expected levels for over half of the days that each sensor recorded data, compared to 99% for adjusted measurements. Therefore, a minority of sensors were consistently recording nighttime concentrations outside the outdoor range. Supplemental Fig. S5 shows raw and adjusted nighttime CO₂ profiles from a sample of sensors within one school on one day for illustrative purposes.

School-day summary statistics

The average adjusted school-day CO₂ concentration in BPS schools in the school year was 841 ppm (SD: 405 ppm). As expected, school buildings with central HVAC had overall lower concentrations and lower variability, averaging 698 ppm (SD: 278 ppm) with CVs ranging from 15% to 53%, compared to buildings with supplemental or no mechanical ventilation with a combined average of 930 ppm (SD: 445 ppm) with CVs ranging from 17% to 63%. One central HVAC school with DCV (henceforth DCV School A) had an average measurement of 531 ppm (SD: 97 ppm) and 99% of observations falling under 1000 ppm, reflecting the successful application of controls. The average classroom-level CV across the school year was 28% but varied widely from 5% to 140%. Supplemental Fig. S6 presents the distributions of adjusted CO₂ measurements stratified by ventilation status, and Supplemental Fig. S7 shows a typical profile of CO₂ levels over the course of a school day (raw and smoothed) for a random classroom. The average CV for a classroom over a single school day was 17% (SD: 9%) but was as high as 155%.

School buildings with central HVAC had the lowest mean concentration during their cooling season (686 ppm; SD: 272 ppm), with slightly higher averages in the shoulder (692 ppm; SD: 273 ppm) and heating (703 ppm; SD: 281 ppm) seasons. Supplemental Table S2 shows similar trends in school buildings without central HVAC comparing the heating season to the non-heating season, albeit with higher seasonal averages and greater mean differences between the seasons. Complete seasonal distributions by mechanical ventilation status are shown in Supplemental Fig. S8.

On average, 14% (SD: 10%) of the variation in one classroom’s CO₂ concentrations could be explained by the variation in concentrations from another classroom in the same building. Buildings with central mechanical ventilation had a slightly higher r² compared to buildings without. Supplemental Fig. S9 visualises correlations between concentrations in a sample of twenty classrooms from a school with central HVAC, with most |r| < 0.3 (representing weak correlations).

In the subsample of complete days (98% of recorded sensor-days), average daily classroom concentrations exceeded 1000 ppm 24% of the time and 2000 ppm 1.3% of the time. Across all classrooms the average maximum school-day CO₂ concentration was 1138 ppm (SD: 562 ppm; 99.9th percentile: 5080 ppm). Maximum concentrations exceeded 1000 ppm on half of the school days and 2000 ppm on seven percent of school days. Fig. 1 shows the distributions of daily mean and maximum classroom CO₂ concentrations for a subset of 20 random schools, stratified by mechanical ventilation system.

Fig. 1.

Distributions of average and maximum daily classroom CO₂ concentrations over a school year in a random subset of 20 schools, stratified by mechanical ventilation system(Central, Supplemental or None). Dashed lines represent CO₂ thresholds from the Boston Public Schools Indoor Air Quality Management Plan (1000 and 2000 ppm).

Classroom CO₂ exceeded 1000 ppm for 24% of school hours on average, and 2000 ppm for 2% of school hours. Table 1 provides more details on summary statistics. Supplemental Fig. S10 shows the distribution of the percentage of each school day where CO₂ concentrations exceeded 1000 ppm for classrooms in the sample of 20 schools. Seasonal trends in daily metrics followed similar patterns as overall summary statistics (Supplemental Table S2).

Table 1.

Summary carbon dioxide (CO₂) statistics by school, classroom and school-day, stratified by mechanical ventilation system.

	All school buildings (n = 125)	Buildings stratified by mechanical ventilation system
		Central (n = 32)	Supplemental (n = 43)	None (n = 50)
	Mean (SD)
Average of all recorded CO2 concentrations (ppm)a	841 (405)	698 (278)	916 (437)	945 (452)
Average of daily maximum CO2 concentrations (ppm)b	1138 (562)	901 (389)	1252 (588)	1309 (613)
Average variability in CO2 measurements relative to mean concentrations (coefficient of variation) (%)a
By school building	38 (10)	32 (10)	41 (8)	40 (9)
By classroom	28 (12)	22 (11)	31 (11)	32 (10)
By school-day	17 (9)	14 (8)	18 (9)	19 (9)
Average of r2 between classroom CO2 levels within school buildings (%)a,c	14 (10)	15 (11)	14 (10)	13 (10)
% of days in school year where daily sensor average CO2 exceeded 1000 ppmb,d	24 (29)	8 (19)	31 (29)	35 (30)
% of days in school year where daily sensor average CO2 exceeded 2000 ppmb,d	1.3 (6)	0.3 (3)	2 (7)	2 (7)
% of days in school year where daily sensor maximum CO2 exceeded 1000 ppmb,d	50 (36)	28 (32)	61 (32)	66 (31)
% of days in school year where daily sensor maximum CO2 exceeded 2000 ppmb,d	7 (16)	2 (8)	9 (17)	11 (19)
% of school hours per school day where CO2 exceeded 1000 ppmb,d	24 (32)	10 (22)	31 (34)	34 (34)
% of school hours per school day where CO2 exceeded 2000 ppmb,d	2 (10)	0.5 (5)	3 (11)	3 (12)

All values in table calculated with the final adjusted dataset for occupied school hours only.

^aCalculated with the full dataset.

^bCalculated with complete days of data (98%).

^cCalculated for each pairing of classrooms within each school building.

^dCO₂ thresholds from the Boston Public Schools Indoor Air Quality Management Plan (1000 and 2000 ppm).

Visualisations and applications

Fig. 2 shows data visualisations co-created with BPS operations staff showing the percent of occupied hours every day where CO₂ exceeded the BPS 1000 ppm threshold in 20 random classrooms of a school with central HVAC (Fig. 2a), and summarised across the school year (Fig. 2b). This figure highlights the spatiotemporal variability in CO₂ levels within one mechanically ventilated school.

Fig. 2.

Percent of a school day with carbon dioxide (CO₂) concentrations exceeding 1000 ppm in 20 random classrooms in a school withcentralmechanicalventilation. Each row represents one classroom, with (a) showing daily values over an entire school year, and (b) showing aggregated yearly percentages. White columns in (a) are weekends and holidays.

Fig. 3 compares daily CO₂ maxima in the prior sample of classrooms with the NVG’s infection-risk-based threshold (Fig. 3a) and the CEN limits for odour control (Fig. 3b) across the school year for illustrative purposes, given the lack of precise information on classroom size and occupancy. Nevertheless, the side-by-side figures demonstrate how different thresholds identify different classrooms and days for interventions.

Fig. 3.

Daily maximum carbon dioxide (CO₂) concentrations in 20 random classrooms in a school withcentralmechanicalventilationcompared to health and comfort-based guidelines. Each row represents one classroom (same one in each figure), with (a) comparing maxima to the Nordic Ventilation Group’s infection-risk-based threshold of 941 ppm and (b) comparing maxima to the European Committee for Standardization’s multi-tiered comfort-based thresholds. White columns are weekends and holidays.

Classroom- and school-specific summary pie charts showing the percent of the school year spent above thresholds (Fig. 4) were found to be the most useful for school operations staff in every-day decision-making and communications. We created this figure iteratively, incorporating feedback from BPS staff until the figures met their needs. More information-rich plots, such as those in Fig. 2, Fig. 3, had less utility for general audiences.

Fig. 4.

Percent of the school year with classroom carbon dioxide (CO₂) concentrations below 1000 ppm, between 1000 and 2000 ppm, and above 2000 ppm in a sample of 12 classrooms inoneschool. Data were restricted to occupied school hours, and charts are ranked by increasing proportion of the school year with concentrations above 1000 ppm.

Discussion

Leveraging a groundbreaking network of 3659 classroom sensors, we mapped the landscape of CO₂ concentrations in a large urban school district and co-created decision tools in what is, to our knowledge, the first academic-school partnership of its kind. Results are generalisable to school districts worldwide with similar ageing infrastructure, and our protocols and code are publicly available. The BU-BPS collaboration illustrates how raw data from commercial large-scale indoor environmental quality sensor networks can be converted into analytical datasets for both school decision-making and research.

This research supported BPS Facilities Management in leveraging comprehensive monitoring data to improve IAQ by identifying and addressing cases of elevated CO₂ and evaluating the impact of interventions.²⁶ Broadly, the BU team has helped BPS detect and correct issues with the sensors and public-facing dashboard, advocate for investments in HVAC systems, develop materials for communications and grant applications, enforce school health policies, and share knowledge with tangible endpoints (e.g., updates to the IAQ Management Plan).²⁶ Supplemental Table S3 summarises additional details. As we show here, simply having access to raw sensor data is not enough to extract the full potential of sensor data, and as we’ve previously published, research based on environmental sensor data is only useful with meaningful stakeholder engagement and action-oriented insights.²⁷ Engaging school district operations staff and aligning with district priorities were key to this project’s success.

This study was novel in characterising CO₂ concentrations for thousands of classrooms over a large school district and an entire school year. Previous studies in schools sampled over shorter time scales (hours to months),^12,28 and the few year-long campaigns29, 30, 31, 32, 33 were neither school-wide nor district-wide. During school hours, CO₂ concentrations in our study ranged from outdoor levels to more than an order of magnitude higher even within one day. Daily average and maximum concentrations in BPS schools largely fell within the range of concentrations reported in other schools. In a review article of schools worldwide, Fisk 2017 reported maximum concentrations ranging from over 1000 ppm to 6000 ppm and average concentrations ranging from outdoor levels to over 5000 ppm.¹² Eichholtz et al., 2023 reported daily averages ranging from 818 to 1000 ppm in a four-year survey of 252 classrooms across 27 primary schools in the Netherlands.²⁹ In a year-long sampling campaign of ten classrooms across five schools in Australia, Andamon et al., 2023 documented daily averages ranging from 657 to 2235 ppm and daily maxima ranging from 1417 ppm to over 5000 ppm.³⁰ Finally, Iyengar et al., 2024 reported monthly average concentrations in a New York City private school with central HVAC ranging from 678 to 817 ppm in SY21-22.³¹ This range was comparable to that of mechanically ventilated BPS schools (results not shown).

We found that regardless of ventilation system, classroom CO₂ concentrations are weakly correlated (most |r|<0.3) within a building and fluctuate substantially over the course of a school day, highlighting the importance and added value of monitoring every classroom. Variability across a school year was typically higher, with greater variation comparing between schools vs. within schools (Table 1). This has implications for epidemiological study design, since these levels of fluctuations suggest potential misclassification in studies relying on spot check data. Given the observed variability, we recommend that at least five CO₂ sensors should be installed irrespective of school size based on a sensor accuracy rating of 10% and assuming randomly selected and well-mixed classrooms. As shown in Supplemental Table S1, the larger the school, the more sensors are necessary to capture unbiased mean estimates (e.g., more than 25 sensors are recommended for buildings with over 75 classrooms). The variability in the BPS CO₂ concentrations has been observed elsewhere^12,28 and reflects the influence of individual building features, ventilation conditions, outdoor CO₂ concentrations, seasonal and other outdoor weather conditions, and occupant characteristics and behaviour.³⁴

Although concentrations in mechanically ventilated schools were typically lower and less variable than in schools without mechanical ventilation (Table 1), there were threshold exceedances for buildings with all three types of ventilation systems (Fig. 1, Fig. 2). Two classrooms in the same mechanically ventilated school with the same number of occupants may have different CO₂ concentrations for many reasons, including differing occupant density and/or behaviour; differing levels of air leakiness; dampers, windows, and doors being open vs. blocked or closed; and ventilation systems working vs. malfunctioning. BPS schools with limited supplemental ventilation had slightly lower concentrations and fewer recorded exceedances compared to schools without mechanical ventilation (Table 1), but their CO₂ distributions largely overlapped (Supplemental Fig. S6). More information is needed to understand the usage of unit ventilators in these buildings, but these results suggest inconsistent use. As expected, schools with working DCV tied to CO₂ concentrations rarely exceeded thresholds. Linking unit ventilators to DCV can enhance ventilation in buildings without central HVAC. Overall, our findings are consistent with prior work which has shown higher VRs in mechanically ventilated classrooms compared to naturally ventilated classrooms.²⁸ Nevertheless, natural ventilation practises such as opening windows and doors can serve as a useful adjunct in school buildings that are not yet equipped with mechanical ventilation.³⁵

Concentrations in BPS schools were highest during the heating season irrespective of ventilation status. During the heating season, occupants may be more likely to close windows, decreasing natural ventilation. BPS schools with central mechanical ventilation have fresh air circulating year-round, though the proportion of fresh air (vs. recirculated air) is generally lower during the heating season. Natural ventilation practices may be more common in schools without mechanical systems outside the heating season, and building leakiness increases with window AC installation. Notably, differences across seasons were smaller than differences across ventilation systems (Supplemental Table S2), with the caveat that BPS’s heating season comprises the vast majority (71%) of the school year. Further work is needed to assess these relative contributions to CO₂ and VRs.

There are few evidence-based CO₂ guidelines for protecting health and learning²⁰ and school staff have expressed the critical need for standards.⁸ CO₂ concentrations are not always well-correlated with concentrations of other airborne pollutants¹⁴ and should not be interpreted as directly representing overall IAQ. Thresholds should be tailored to different space uses and characteristics, and specific endpoints of interest.^14,20 We demonstrate several threshold-based CO₂ metrics that encompass exposure duration and intensity, and with utility for different applications (Fig. 2, Fig. 3). Numerous guidelines specify daily maxima corresponding to desirable VRs. Continuous monitoring is necessary to capture those values (although recorded daily maxima can still be underestimates²⁰ or overestimates that reflect transient events; smoothing minute-level measurements can reduce the occurrence of false spikes).³⁶ Continuous CO₂ data can also be used to derive VRs (as demonstrated in Yuan et al., 2025)¹⁹ and ensure compliance with evidence-based VR guidelines.³⁷

Limitations of our data included data missingness, noise, residual calibration error, and sensor drift, although we developed detailed data cleaning and adjustment protocols. Reasons for missing data included sensor malfunction, power outages, and manual disconnecting of power sources, but most sensors had high completeness in recorded data (>90%). We identified minute-to-minute fluctuations and outliers in the raw data and developed protocols to correct for noise and outliers. We used offset adjustments based on nighttime values to minimise residual calibration bias. Ten percent of sensors had consistently large offsets but were retained for this study since their random distribution across schools made it unlikely for the overall characterisations to be biased. There is a potential for offset overadjustment if CO₂ levels do not equilibrate with nighttime concentrations, which could happen at very low AERs (e.g., 0.1 h⁻¹). However, our previous work showed that the mean AER in BPS classrooms was 3 h⁻¹ and even though the minimum AER recorded was 0.1 h⁻¹,¹⁹ the corresponding overadjustment is negligible. Longitudinal drift was identified in a subset of sensors but was not adjusted given the small magnitude (<100 ppm over the school year for most sensors). Non-dispersive infrared sensors are considered stable and durable but can be affected by factors such as temperature and length of use.³⁴ In future studies, data from sensors with large offsets and/or drift and classrooms with low AERs may require further processing or exclusion. Despite these limitations, preliminary findings from our validation studies, which demonstrate consistency between the classroom sensors and validation sensors, lend credibility to observed trends. Inevitably there are trade-offs when using commercial sensors that are deployed on a large scale in schools,³⁸ justifying the need for partnerships and the transparent sharing of protocols. A lack of detailed information on classroom sizes, daily break schedules, and occupants limited our ability to stratify results and assess drivers of variability. Future work will explore these factors in depth. Nevertheless, the primary strength of this study lies in the comprehensive characterisation of CO₂ concentrations across an entire school district over a school year, which has directly facilitated IAQ interventions.

Our academic-school partnership has enabled actionable research in support of IAQ improvements. The heterogeneity of CO₂ levels across school classrooms and between schools underscores the need for continuous classroom-level monitoring and demand-controlled ventilation. Results and protocols are generalisable to other schools and researchers worldwide. Our work demonstrates how expanding the IAQ analytic capacity of schools empowers both effective interventions and scientific advancements, which will result in healthier school environments for students and staff.

Contributors

BG, KT, MPBM, KHW, and MPF conceptualised the study. BG, KT, PV, MY, and MPF designed study methodology. KHW, BT, LB, LM, and MPF acquired study data and resources. MPF obtained funding. KT and PV were responsible for data curation. BG, KT, PV, and YY conducted formal analysis and visualisation of the data using software. BG, PV, and YY performed data validation. BG and KT conducted project administration and BG, MY, and MPF supervised the study. BG, KT, and MPF drafted the original manuscript. BG, KT, PV, MPBM, YY, KHW, LB, LM, MY, and MPF reviewed and edited the manuscript. All authors had full access to all the data in the study and accept responsibility for the decision to submit the manuscript for publication.

Data sharing statement

The minute-level CO2 data used in this study are not publicly available.

Declaration of interests

KHW received approved support for conference travel from the sensor vendor, SGS Galson.