Evaluating Indoor Air Quality Controls Under Occupancy Uncertainty: A Conceptual Framework for Exposure Inference, Uncertainty Quantification, and Intervention Assessment

1. Introduction

Indoor air quality (IAQ) evaluation often treats the monitored room as the relevant unit of analysis. That simplification is practical when occupants remain stationary or when the purpose is compliance with room-based concentration thresholds. It is less satisfactory when controls are intended to reduce personal exposure across multiple spaces, because exposure depends on movement between rooms, unequal occupancy durations, and uncertainty in those occupancy assignments. The problem is especially important in schools, childcare settings, homes, and public buildings, where pollutant sources, ventilation, and occupancy vary over time and where susceptible groups may be present. [1] [2] [3]

Available studies provide pieces of the required evidence but do not yet offer a unified framework for evaluating IAQ controls under occupancy uncertainty. Filtration interventions in child-serving facilities show that enhanced filtration can reduce indoor particulate matter, yet effectiveness varies across standard schools, nurseries, childcare houses, portable buildings, and large rooms, indicating that room type and air-mixing assumptions matter for causal interpretation. Multi-zone smart-building monitoring demonstrates that IAQ conditions differ across functional zones and that event-driven sensing can reduce computational burden while retaining responsiveness to meaningful environmental changes. Occupant-centered assessment based on timetables shows that room sensors can be linked to inferred occupant locations without wearable devices. Digital home monitoring studies show that long-duration in-home data collection is feasible but also reveal constraints related to usability, affordability, and the limited environmental determinants captured by a monitor placed in one room. [1] [4] [2] [3]

This article addresses a narrower methodological question than general building control design: how should IAQ controls be evaluated when occupancy is uncertain? The proposed answer is a conceptual framework that integrates room measurements, probabilistic occupancy inference, exposure aggregation, and uncertainty quantification. The aim is not to report new empirical results. It is to specify operational definitions, estimands, algorithms, and validation procedures that can support future intervention studies in occupancy-uncertain settings.

2. Research question and evidence scope

The central research question is: how can the effect of an IAQ control be evaluated on occupant exposure when the monitored environment is multi-zone and the occupants’ exact locations are only partially observed? This question has three subproblems: first, inference of time-varying exposure from room-level measurements; second, quantification of uncertainty arising from unknown or imperfectly known occupancy; and third, assessment of intervention effects in a way that distinguishes room-level pollutant reduction from occupant-level exposure reduction.

The evidence base used here is a targeted narrative synthesis rather than a systematic review. The synthesis is bounded to six studies that directly inform the problem formulation. One study evaluated enhanced filtration interventions in child-serving facilities and found that particulate matter reduction occurred in standard schools and a nursery but was inconsistent in childcare houses, portable buildings, and very large rooms; that work also used indoor/outdoor ratios and single-zone mass balance models to interpret intervention performance. A second study developed multi-zone, multi-parameter IAQ monitoring using event-driven neuromorphic sensing and documented zone heterogeneity together with adaptive filtering of temporally noisy signals, including occupancy-linked CO2 and VOC fluctuations. A third study shifted assessment from rooms to persons by using timetable-based location inference and per-person aggregation of air quality and thermal comfort. A fourth study examined the feasibility of digital IAQ monitoring in homes, showing sustained data capture was possible but also noting limits of a bedroom-only monitor and absence of real-time feedback. Two additional studies provide transferable methods: one maps IAQ machine-learning tasks to data structures and stresses preprocessing, missing data, and non-stationarity; the other illustrates uncertainty quantification using conformal prediction and distinguishes useful predictive intervals from misleadingly narrow ones. [1] [4] [2] [3] [5] [6]

The scope therefore excludes unrelated building systems and does not attempt a general theory of all occupancy sensing. The focus is evaluation of IAQ controls—especially filtration and monitoring-supported control assessment—under occupancy uncertainty, with emphasis on exposure inference, quantified uncertainty, and intervention-effect estimation.

3. What prior work establishes and what remains unresolved

Existing intervention evidence establishes that room-level IAQ improvements do not translate uniformly across building types. In child-serving facilities, enhanced filtration reduced PM2.5 and PM10 in standard schools and a nursery, and a linear mixed-effects model indicated PM2.5 reduction by portable air cleaners, but predictive consistency from single-zone mass balance models deteriorated in portable buildings, very large rooms, and childcare houses. High indoor/outdoor PM2.5 ratios in childcare houses suggested important indoor sources such as cooking and re-entrained dust. These findings imply that evaluation frameworks must account for structural heterogeneity and indoor source variability rather than assuming a common response to the same intervention. [1]

Monitoring studies establish that IAQ is inherently multi-parameter and non-stationary. Multi-zone monitoring work in a smart educational building sampled PM, CO2, CO, TVOCs, and O3 across nine distinct zones and reported that adaptive Kalman filtering reduced stochastic noise while preserving meaningful environmental changes. The same study noted substantial temporal noise in some CO2 and VOC signals due to changing occupancy. This is relevant because occupancy uncertainty is not merely a missing-data nuisance; it is a generator of observable concentration dynamics and a source of ambiguity in attribution. [4]

Occupant-centered assessment work resolves one important limitation of room-based evaluation by showing that a scheduling file can dynamically associate individuals with rooms through the day. That approach is attractive where timetables are institutionalized, as in schools and universities, because it avoids wearables and dedicated tracking hardware. It also shows that exposure can be aggregated per participant and per day. Yet the method, as reported, presumes the timetable is sufficiently reliable for location inference. It does not directly model uncertainty from absences, unscheduled movement, timetable deviations, or non-institutional settings such as homes. [2]

Home monitoring feasibility studies clarify a different boundary condition. Monitors can be installed for extended periods and can collect continuous data from a child’s bedroom, but participant usability, affordability, and internet access affect real-world deployment. The monitor in that study also could not detect indoor allergens such as mold or dander, and the absence of real-time data limited behavior change during monitoring. For occupancy-uncertain evaluation, the lesson is that feasible monitoring does not guarantee that the measured room represents total exposure, nor that all relevant hazards are observed. [3]

Methodological reviews of IAQ machine learning add two requirements. First, IAQ data are heterogeneous time series with missing data and non-stationary sources, so preprocessing choices are consequential. Second, there is no universal best algorithm; the model should match the task and dataset structure. Transferable uncertainty work adds a third requirement: predictive intervals must be judged by calibration as well as width, because narrow intervals can be practically useless if they are overconfident. Together these studies motivate an evaluation design that explicitly represents occupancy uncertainty instead of embedding it silently in model residuals. [5] [6]

4. Conceptual model and operational definitions

The proposed framework distinguishes four entities: zones, occupants, controls, and observation intervals. Let z denote a zone, i an occupant, c a control condition, and t a discrete time interval. A zone is any monitored indoor space with a defined sensor location or a validated imputation rule for that location. An occupant is any person for whom exposure is to be estimated. A control condition is the state of the evaluated intervention, such as baseline filtration versus enhanced filtration, or control-off versus control-on periods. Time intervals should align with the native sensor resolution or a justified aggregation interval.

Room-level pollutant concentration is written as C(z,t,p), where p indexes the pollutant or thermal variable. Occupancy probability is written as P(i,z,t), interpreted as the probability that occupant i is in zone z during interval t conditional on available information. The exact operational rule is: for every occupant i and interval t, sum over all eligible zones must equal 1, that is, Σ_z P(i,z,t) = 1. If an occupant is known to be absent from the building, the zone set includes an explicit outside or unobserved state so that normalization is preserved.

Expected interval exposure for occupant i to pollutant p at time t is defined as E(i,t,p) = Σ_z P(i,z,t) × C(z,t,p). Cumulative expected exposure over an evaluation window T is CE(i,p,T) = Σ_t w(t) × E(i,t,p), where w(t) is the interval duration in hours or another consistent unit. If the purpose is threshold exceedance burden rather than cumulative dose, the outcome may instead be defined as time above threshold: TE(i,p,T; tau) = Σ_t w(t) × I[E(i,t,p) > tau], where I[·] is an indicator and tau is a pollutant-specific hypothetical threshold selected a priori.

Intervention effect can then be defined at two levels. The room-level effect for zone z is Delta_room(z,p) = mean_t C_on(z,t,p) minus mean_t C_off(z,t,p). The occupant-level effect is Delta_occ(i,p) = CE_on(i,p,T) minus CE_off(i,p,T). These estimands need not agree. A control can improve average room conditions while leaving some individuals with little benefit if they spend time in poorly controlled zones or if occupancy shifts after implementation. This separation is the main conceptual gain of the framework.

5. Proposed methodology: occupancy inference and uncertainty quantification

The framework uses a layered inference strategy because occupancy information varies by setting. Tier 1 is deterministic schedule assignment. If an institutional timetable exists, each occupant is assigned to the scheduled zone for each interval. Tier 2 introduces probabilistic deviations from the schedule, such as late arrival, absence, temporary relocation, or corridor time. Tier 3 allows data-informed adjustment using IAQ-related signals or auxiliary building data when available, such as motion, door events, ventilation state, or occupancy-sensitive CO2 patterns. The tiers can be applied sequentially, with higher tiers activated only when data support them. [2] [5] [4]

A practical decision sequence is as follows. Step 1: define the zone graph and sensor-to-zone mapping. Step 2: preprocess environmental time series using documented filtering, calibration, synchronization, and missing-data handling rules suited to the task. Step 3: generate an initial occupancy distribution P0(i,z,t) from timetables, room rosters, household routines, or explicit assumptions. Step 4: update P0 to P1 using any direct occupancy evidence. Step 5: if no direct evidence exists, estimate deviations using an occupancy-likelihood model informed by context variables, while constraining transitions to physically plausible movements between zones. Step 6: compute expected exposure E(i,t,p) and cumulative outcomes. Step 7: quantify uncertainty using conformal prediction around either occupancy probabilities, exposure trajectories, or intervention-effect estimates. [2] [5] [6]

Conformal prediction is proposed here because it can provide distribution-free interval guarantees under stated exchangeability conditions and because the transferable lesson from uncertainty research is that reliability depends on calibration, not only on narrowness of intervals. For this IAQ problem, one implementation is split conformal exposure prediction. The calibration target is the absolute residual between observed or proxy-validated exposure and predicted expected exposure. For a chosen miscoverage level alpha, the conformal radius q_(1-alpha) is the empirical (1-alpha) quantile of calibration residuals. The interval for cumulative exposure becomes [CE_hat(i,p,T) - q_(1-alpha), CE_hat(i,p,T) + q_(1-alpha)]. When observed personal exposure is unavailable, calibration can use intervals built around validated occupancy labels in a substudy or around held-out room-transition observations. [6]

The framework deliberately avoids claiming that more complex models are always preferable. IAQ machine-learning evidence indicates that no universal algorithm dominates across all dataset structures. Therefore, occupancy-likelihood models should be selected by task: multinomial logistic regression or hidden-state models for interpretable transition probabilities; tree-based methods when interactions between schedule, time, and context are strong; sequence models only when sufficient longitudinal depth and stable preprocessing are available. Simpler models are preferable when direct occupancy data are sparse, because apparent precision without reliable calibration would be misleading. [5] [6]

A hypothetical operational rule may guide implementation: if the estimated 90% conformal interval for Delta_occ(i,PM2.5) includes zero for more than a prespecified fraction of occupants, the intervention is judged inconclusive at the occupant level even when mean room concentrations decrease. This rule is illustrative rather than validated; its value lies in making the decision logic explicit and falsifiable.

6. Comparative assessment of IAQ controls under occupancy uncertainty

The comparator structure should match the control being evaluated. For filtration interventions, the primary comparison is baseline versus enhanced filtration periods within the same zone network, with outdoor conditions and major indoor source events documented where possible. The child-serving facility study used indoor/outdoor ratios to interpret filtration effectiveness, an approach that should be retained because outdoor infiltration affects indoor particulate levels and can otherwise confound apparent intervention effects. In occupancy-uncertain evaluation, indoor/outdoor ratios remain a room-level diagnostic rather than the final outcome, which should be occupant-weighted exposure reduction. [1]

The proposed framework extends, rather than replaces, single-zone mass balance reasoning. Where room type and airflow conditions make single-zone approximations plausible, mass balance models can generate expected room-level concentration changes under filtration. These predicted changes can then be propagated through occupancy probabilities to generate predicted exposure changes. Where prior evidence suggests the single-zone approximation is unreliable—portable buildings, very large rooms, repurposed houses, or spaces with strong indoor source heterogeneity—the framework marks those zones as structurally uncertain and widens effect interpretation accordingly. [1]

For smart-building monitoring, event-driven sensing contributes mainly to scalable data acquisition, not to exposure inference by itself. The neuromorphic study is relevant because it shows that selective activation can preserve meaningful environmental events across multiple zones while reducing computational load. In an occupancy-uncertain evaluation design, such sensing can support dense spatial coverage without requiring all streams to be processed continuously at full cost. Yet any event-driven representation must still be translated back to concentration estimates or threshold-state estimates compatible with exposure metrics. Energy efficiency is therefore a secondary evaluation criterion, whereas occupant exposure remains primary. [4]

Compared with timetable-based occupant-centered assessment, the present proposal introduces uncertainty explicitly. The timetable method demonstrates that fixed-room sensors can support person-level aggregation without wearable devices, which is an important foundation. The present framework departs from that method by replacing deterministic room assignment with probability distributions over plausible locations and by attaching interval estimates to cumulative exposures and intervention effects. This extension is necessary when attendance is irregular, movement is unscheduled, or the setting lacks formal timetables, such as homes or mixed-use community facilities. [2] [3]

7. Prospective evaluation protocol

Because this manuscript is conceptual, the empirical results section is replaced by a prospective protocol. The protocol is suitable for schools, childcare facilities, offices, or homes with repeated monitoring. The minimum data elements are: zone-level IAQ measurements for target pollutants and thermal variables; time stamps synchronized across zones; intervention state indicators; a source of occupancy information, such as timetables, rosters, routines, or partial direct observations; and metadata on building type, room function, and known source events. For particulate filtration studies, PM2.5 should be mandatory because prior intervention evidence directly concerns that pollutant. CO2 should also be included when ventilation adequacy and occupancy-related emissions are relevant. [1] [2] [5]

The protocol has five phases. Phase A, design: specify the estimand, primary pollutant, comparison windows, and validation substudies before data collection. Phase B, acquisition: collect multi-zone measurements with documented calibration, sensor placement, and data completeness criteria. Sensor placement should be justified relative to the respiratory zone and away from obvious interference where feasible, consistent with prior home monitoring practice. Phase C, occupancy modeling: construct P(i,z,t) using the tiered inference approach and record the assumptions used for each occupant-time segment. Phase D, effect estimation: compute room-level and occupant-level outcomes for baseline and intervention periods, with conformal intervals. Phase E, adjudication: interpret whether the intervention is supported, inconclusive, or unsupported at both room and occupant levels. [3] [6]

Validation should occur at three levels. First, concentration validity: compare processed sensor outputs with reference instruments or previously validated devices where available. Second, occupancy validity: for a purposive subset of intervals, compare inferred occupancy with direct observations, access logs, or ethically acceptable short-duration tracking. Third, exposure validity: where personal monitors or trusted proxies are available for a subset, compare estimated exposures CE_hat with observed exposures. The evaluation metrics should include calibration of prediction intervals, mean interval width, bias in cumulative exposure, and classification performance for threshold exceedance burden. Transferable uncertainty work suggests evaluating coverage jointly with interval width, because narrow but miscalibrated intervals create false confidence. [3] [6]

A concrete decision sequence for analysis is as follows: (1) preprocess and synchronize all time series; (2) label intervention periods and censor intervals dominated by extraordinary events if prespecified; (3) estimate occupancy probabilities; (4) compute room concentrations and occupant exposures; (5) construct conformal intervals; (6) compare Delta_room and Delta_occ; (7) perform sensitivity analyses under alternative occupancy priors; and (8) report failure cases where conclusions depend materially on unverifiable assumptions. This sequence makes occupancy uncertainty a reported component of the study rather than an invisible source of residual variance. [5] [6]

8. Assumptions, failure cases, and limits of generalization

The framework rests on several assumptions. It assumes that the monitored variables are relevant proxies for the intervention target and for exposure in the studied setting. It assumes that occupancy uncertainty can be represented as probabilities over a finite zone set. It assumes that measurement timing and zone definitions are sufficiently stable for exposure aggregation. For conformal prediction, it assumes that the calibration regime is comparable to the application regime; strong distribution shift would weaken coverage guarantees. [6]

Several failure cases should be anticipated. First, severe within-room heterogeneity can invalidate the use of one sensor as a zone representative, especially in large rooms or complex airflow settings. This concern is consistent with inconsistent modeling performance in portable buildings, childcare houses, and very large rooms. Second, important indoor sources can dominate filtration effects, as indicated by elevated indoor/outdoor ratios in some childcare houses. Third, timetable-based or routine-based occupancy inference may fail during irregular schedules, shared spaces, or informal movement patterns. Fourth, home monitoring from a bedroom may miss exposure elsewhere in the dwelling and cannot capture hazards not sensed by the monitor, such as allergens outside the device’s scope. Fifth, event-driven sensing can compress data efficiently but may complicate reconstruction of conventional exposure metrics if calibration between spikes and concentrations is not maintained. [1] [2] [3] [4]

Generalization should therefore be bounded. The proposal is most suitable for multi-zone environments with fixed or semi-fixed occupants, repeated schedules or routines, and interventions whose plausible mechanism acts through measured pollutants. It is less suitable for highly mobile populations without any occupancy anchors, for exposures dominated by unmonitored hazards, or for buildings where air-mixing behavior changes abruptly and unpredictably. The framework also does not establish health benefit directly; it evaluates pollutant exposure reduction under uncertainty, which is one step in a larger causal chain.

9. Conclusion

This article has established, at a conceptual level, that evaluating IAQ controls under occupancy uncertainty requires a shift from room-only concentration changes to occupant-weighted exposure effects with explicit uncertainty bounds. The targeted synthesis supports four premises: filtration can improve particulate conditions but not uniformly across building types; multi-zone sensing is necessary because IAQ is spatially heterogeneous; timetable or routine information can support person-level assessment without wearables; and uncertainty quantification must balance interval width with calibration rather than equating narrowness with reliability. [1] [4] [2] [6]

The proposed contribution is a bounded methodological framework for this specific problem. It defines variables, exact normalization and exposure rules, layered occupancy inference, occupant-level and room-level estimands, and a prospective validation protocol. It also specifies how prior methods can be combined and where they are insufficient when occupancy is uncertain. What remains untested is the empirical performance of the framework across real intervention datasets, including the calibration of conformal exposure intervals, the comparative value of different occupancy models, and the robustness of conclusions in settings with irregular movement or strong indoor sources. Those questions require future implementation studies and should be treated as open empirical tests rather than as established findings. [5] [6]

References

1. Xia, Tian; Raneses, Julia; Locke, Morgan; Garcia, Raquel; Walding, Alison; DeMajo, Ricardo F.; Schulz, Amy J.; Batterman, Stuart A. (2026). Effectiveness of enhanced filtration interventions for improving indoor air quality in child-serving facilities. Frontiers in Pediatrics. 14, 1875263. doi:10.3389/fped.2026.1875263.
2. Ruffa, Filippo; Iadarola, Grazia; De Capua, Alberto; De Capua, Claudio (2026). A Method for Optimized Monitoring of Indoor Air Quality in Public Buildings. Sensors. 26(14), 4559. doi:10.3390/s26144559.
3. Orr, Olivia; Honegger, Kyle; Sanchez, Lizbeth; Bhat, Kiran; Nimmagadda, Sai; Kan, Kristin (2026). Feasibility and Acceptability of a Digital Indoor Air Quality Monitor Among Children with Asthma. JMIR Pediatrics and Parenting. 9, e92070-e92070. doi:10.2196/92070.
4. De Jesus, Luigi Carlo M.; Africa, Aaron Don M.; Illahi, Ana Antoniette C.; Gustilo, Reggie C.; Brucal, Stanley Glenn E. (2026). Spike-Driven Neuromorphic Sensing for Energy-Proportional Indoor Air Quality Monitoring in Multi-Zone IoT-Enabled Smart Building Environments. Sensors. 26(13), 3992. doi:10.3390/s26133992.
5. Rosca, Cosmina-Mihaela; Stancu, Adrian (2026). Machine Learning Applications with Sensors for Indoor Air Quality Research. Sensors. 26(9), 2909. doi:10.3390/s26092909.
6. Abdi, Erfan; Sattari, Mohammad Taghi; Pal, Mahesh; Milewski, Adam; Apaydin, Halit (2026). Uncertainty Estimation in Predicting River Discharge Using Probabilistic Machine Learning and Conformal Prediction. Sensors. 26(18), 5800. doi:10.3390/s26185800.