Pulmonary function testing (PFT) provides physiological measurements of the respiratory system across multiple dimensions, typically classified into (1) spirometry, which measures air flow, lung volumes, and capacities during a expiratory forced vital capacity (FVC) maneuver; (2) static lung volumes; and (3) gas exchange parameters such as the diffusing capacity for carbon monoxide and oxygen saturations [1]. PFTs are critical for the diagnosis and prognostication of respiratory disorders, and provide a noninvasive method for measuring and monitoring the degree of respiratory impairment [2]. They are recommended for the initial evaluation of patients with chronic dyspnea and other respiratory symptoms, as well as for individuals at risk of respiratory complications due to transplant or surgery [3,4].

Basic spirometry remains the most widely used component of PFT, largely due to its size and portability, allowing it to be performed in clinic office settings or remotely at home with adequate training. However, spirometry, by definition is an expiratory FVC maneuver that focuses on assessing airflow limitations and does not directly measure static lung volumes, which can be integral to understanding many respiratory conditions [4]. Accurate determination of static lung volumes traditionally necessitates more complex and resource-intensive techniques such as body plethysmography or gas dilution methods, with body plethysmography serving as the current gold standard [3,5,6]. However, these methods, while precise, may not always be readily accessible, cost-effective, or suitable for routine clinical practice outside a specialized pulmonary function laboratory.

Advancements in artificial intelligence (AI) techniques have introduced new avenues in health care, offering the potential to derive comprehensive insights from existing data, including patterns not easily recognizable through human interpretation or standard statistical modeling. A prior study by Beverin et al [7] examined the prediction of total lung capacity from spirometry using three tree-based machine learning (ML) models, achieving a mean squared error of 560.1 mL. They further developed models to classify restrictive ventilatory impairment, achieving a sensitivity and specificity of 83% and 92%, respectively. However, they did not explore prediction of the complete lung volume assessments. Predicting functional residual capacity status, for example, could facilitate the prevention of atelectasis during anesthesia [8]. Another study by Evankovich et al [9] developed a regression model in patients with chronic obstructive pulmonary disease (COPD) to predict residual volume and its elevation status, achieving an area under the receiver operating characteristic curve (ROC) of 0.95 for predicting residual volume above 175%. However, these models lack applicability beyond the COPD cohort [9]. Given this context, we hypothesized that ML models could predict static lung volumes using spirometry alone across a diverse cohort of lung conditions. Such an approach could reduce the need for identifying those who would benefit most from formal lung volume assessments. In this study, we applied ML approaches to develop and validate an algorithm for estimating lung volumes and capacities from standard spirometry. We further examined the model performance among subsets of physiologic derangements such as obstructive and restrictive ventilatory disorders.

A total of 121,498 PFTs were used in this study, with 85,017 allocated for exploratory data analysis and model development and 36,481 tests reserved for model evaluation. The median age across the cohort was 64.7 years (IQR 18‐119.6), with a nearly balanced gender distribution between genders, with 48.2% (n=58,607) female patients and 51.8% (n=62,889) male patients. The cohort was predominantly White (n= 114,388, 94.1%), followed by African American patients (n=4,656, 3.8%). Of particular importance, the distribution of baseline PFT measures—both spirometry and lung volumes—showed no differences between the training and testing datasets. Standardized mean differences, indicating the degree of difference between the training and testing sets, were minimal across all variables, suggesting a well-balanced model development and testing cohorts. A complete breakdown is provided in Table 1.

Multimedia Appendix 3 stratifies the same cohort according to the ATS classification criteria for pulmonary function patterns (ie, normal, obstructive, restrictive, and mixed pattern). This stratification highlights differences in demographics and pulmonary function measures between individuals with normal, obstructive, restrictive, or mixed patterns assigned using spirometry. Predictably, spirometry measures—including FEV1, FVC, and the FEV1/FVC ratio—significantly differed between groups (P values<.001), as did all phenotype-related parameters presented in the table.

The development of ML models to predict lung volume status (normal vs abnormal findings) from spirometry in over 110,000 patients has yielded highly encouraging results, displaying remarkable discriminatory power with high AUC values (0.81‐0.95) across measured lung volumes. Estimates of FRC, TLC, RV, and the RV/TLC ratio status show strong sensitivity and specificity. These metrics remain largely consistent across spirometry-defined pattern subgroups, with a few exceptions that can generally be attributed to the rarity of abnormal lung volume measures in certain spirometry patterns. The ability to predict lung volume measures without having to perform extensive testing represents a promising innovation for improving the diagnosis and management of dyspnea and chronic respiratory diseases, particularly in the primary care setting [16]. The strong predictive performance of lung volume measurement underscores the potential of these models as a transformative tool in respiratory medicine, offering substantial clinical implications and opportunities for enhancing patient care.

The performance of the regression models showed a high correlation between the training and testing datasets, suggesting that the models were able to effectively capture the relationship between spirometry-derived features and measured lung volumes and capacities derived from body plethysmography. The effectiveness of the models was evident in their ability to closely approximate lung volumes with minimal deviation from true values on average. The RMSE and MAE values are low relative to their respective lung volume ranges. For instance, the median TLC measure in the cohort was 5.5 L, with the model attaining an MAE of 0.43 L and an MAPE of 7.92%. The ability to accurately estimate the RV/TLC ratio further highlights the potential of these models in capturing the dynamic interplay between these volumes, which is particularly relevant in differentiating between common lung conditions such as COPD, asthma, and restrictive lung diseases [17-20]. The high R-squared values observed for TLC (0.87 in the training set and 0.85 in the testing set) underscore the model’s capacity to capture a significant portion of the variance in TLC measurement. Similarly, the robust estimation of RV (R-squared of 0.73 in the training set and 0.71 in the testing set) and FRC (R-squared of 0.78 in the training set and 0.75 in the testing set) further validates model reliability in estimating lung volumes crucial for the evaluation of respiratory function. The model demonstrated a high correlation for vital capacity (R²=0.98). However, this finding is misleading, as spirometry already provides an accurate estimate of vital capacity, making it trivial to map to a similar value obtained via body plethysmography, assuming minimal measurement error and consistent effort on the part of the patient when executing breathing maneuvers. A significant change in TLC has been reported to be 10% over one year, whereas this model was able to predict TLC within 7.5% and 550 mL [10]. No significant changes were reported in FRC or RV over time. Considering the performance metrics as a whole, the potential of these models to augment clinical practice is encouraging, with R-squared values exceeding 0.7 for all volumes except ERV, which seems to be the most challenging volume to predict accurately. Estimation of TLC, RV, and their ratio (RV/TLC) is particularly promising, as the accurate estimation of the RV/TLC ratio facilitates the identification of air trapping and hyperinflation, which are key factors in many patients’ symptomatology [3,17-20]. Moreover, the reasonable estimation of FRC suggests its potential utility as an indicator for restrictive lung disease diagnosis and treatment. This is particularly important as body plethysmography directly measures only FRC, which is then used to calculate the other variables.

Focusing on the estimation of ERV, the notably high MAPE indicates a relatively subpar overall performance. Given that ERV has the narrowest range of measured values (ie, median 0.8 L, (IQR 0-44) L and a large RMSE of 0.31 relative to the ERV range, this elevated MAPE may be partially influenced by the smaller margin for error [21]. ERV measures the volume of air that an individual can exhale after completing a normal tidal breath. Pairing this with spirometry, individuals with a higher ERV may experience more difficulty with exhalation or exhibit an obstructive pattern on spirometry with a lower FEV1 measure [22,23]. A higher ERV could be a sign of lung hyperinflation, while other factors like obesity, pregnancy, and significant ascites can decrease ERV [22,24]. Lung hyperinflation in obstructed patients, which is defined as elevated FRC, RV, RV/TLC, or occasionally ERV, is highly variable in patients and occurs inconsistently over time [23,25]. This inconsistency, combined with ERV’s narrow range, makes it challenging to predict.

Highlighting a more robust model, predictions for the RV/TLC ratio are strong overall, with AUC values ranging from 0.8 to 0.86 across all patterns and 0.91 in the full cohort. Except for normal pattern PFTs, the model consistently achieved sensitivities >0.84, but it struggled to identify positive cases in normal spirometry tests. While spirometry alone does not directly measure RV or TLC, FEV1 and FVC can indirectly reflect changes in lung volumes. In obstructive lung diseases, a reduction in FEV1/FVC ratio combined with an increase in the RV/TLC ratio often indicates air trapping [22-25]. In restrictive diseases, such as pulmonary fibrosis, spirometry may show decreased FVC with a preserved or decreased RV/TLC ratio, suggesting reduced air trapping [22-25]. Given the absence of abnormal FEV1 and FVC values, normal spirometry patterns would not usually suggest the existence of an abnormal RV/TLC ratio, potentially explaining the reduced sensitivity to predicting abnormal RV/TLC in normal spirometry.

A previous study used a CatBoost model to predict the TLC from spirometry, yielding good results [7]. The study reports an MSE of 560.1 mL for TLC and a positive predictive value for reduced TLC of 8% or 67%, depending on the model parameters. However, this study only focused on TLC and did not assess other pulmonary physiologic parameters obtained through lung volume measurements, such as FRC and RV. These parameters are necessary as they are crucial for assessing prognosis in various respiratory diseases [26-30].

Several studies have highlighted the importance of lung volume assessments for the diagnosis and prognosis of respiratory diseases [31]. In routine practice, it can aid in the early detection, diagnosis, and monitoring of respiratory conditions such as COPD, restrictive lung diseases, and neuromuscular disorders affecting respiratory function [10,32,33]. For instance, lung volume measurements (specifically, FRC and TLC) strongly correlate with mortality risk among patients with idiopathic pulmonary fibrosis [27,28,30]. This illustrates that the prediction of lung volumes from traditional spirometry holds substantial promise in clinical scenarios where lung volume measurements cannot be directly performed, such as primary care offices, or health care facilities in rural areas where the equipment for measuring lung volumes is not readily accessible. Another scenario is when a patient is not capable of physically performing lung volume measurements, which could involve physical conditions that prevent them or any number of other limitations that could potentially limit them. Additionally, it may facilitate personalized treatment plans by providing a more nuanced understanding of a patient’s lung capacities, as lung volume measurements are typically performed only after a patient is determined to have an abnormal spirometry, unless in specialized centers.

Accurate assessment of lung volumes is pivotal in diagnosing and monitoring various respiratory conditions, including COPD, interstitial lung diseases, neuromuscular disorders, and restrictive lung diseases [4,32]. If lung volume measurements are not performed, vital capacity is often used as a surrogate [34,35]. However, there is a significant error in the application of this method, as a reduced vital capacity can be seen in restrictive lung disease and obstructive lung disease with increased residual volume [36]. A restrictive defect on lung volume measurements has rarely been seen occurring with normal vital capacity, and approximately 58% of the time with low vital capacity measurements [36]. Another study showed that when forced vital capacity >100% predicted in males or >85% predicted in females ruled out a restrictive pattern on lung volumes [37]. The use of direct lung volume prediction models, such as those developed in this study, have a significantly better performance than those used in these prior studies and could reduce the frequency of clinical scenarios where lung volumes are unknown.

The AI model’s ability to estimate lung volumes from readily available spirometry data streamlines these diagnostic procedures. A typical spirometry test may take approximately 30‐45 minutes, while lung volume measurements add another 15‐30 minutes [38,39]. Replacing or complementing traditional, more resource-intensive lung volume measurement techniques with the AI model’s predictions from spirometry data offers cost-effective alternatives. The physician fee for spirometry ranges from $29.62 to $150.68, depending upon the medications used, while measuring lung volumes adds another $59.98 to the cost [40]. This approach optimizes healthcare resources, reduces patient burden associated with additional tests, and potentially increases the efficiency of healthcare delivery.

The accessibility of spirometry in various healthcare settings, coupled with the estimation of both lung volumes via the developed models, opens avenues for telemedicine applications. Remote monitoring and assessment of spirometry are already being performed and could be facilitated and enhanced with automated decision support systems utilizing models such as those developed in this study [41-43]. Such strategies could enable the continuous monitoring of patients with chronic respiratory conditions that affect lung volumes [41-43]. This aligns with the evolving landscape of telemedicine, emphasizing its potential in respiratory care.

Despite the remarkable performance of the predictive models, certain limitations warrant consideration. Model training and testing relied on datasets with potential biases in demographic variables, including a majority-White population (91%) of older adults (median age 64.7) years. These factors potentially limit the generalizability to diverse populations, although this model was developed with patients of all ages from two distinct regions of the United States (Midwest and Southeast). Further validation across broader demographic groups from various clinical settings is essential to establish widespread applicability and reliability. Moreover, continuous refinement and validation of the models using larger datasets encompassing a broader spectrum of respiratory conditions and disease severities is imperative. This iterative process would enhance model performance while preventing model drift, ensuring its efficacy in diverse clinical scenarios even as standard clinical practices are updated or changed.

In conclusion, the development of AI models for predicting lung volumes from spirometry represents an advancement in pulmonary function assessment. The remarkable sensitivity and specificity offered by the classification models affect a transformative approach to complement traditional lung volume measurement techniques. While the regression models may not attain the same level of performance, the continuous nature of their estimates provides a unique addition to supplement and contextualize binary classifications, potentially elucidating new insights into the remote monitoring of pulmonary function. If integrated into clinical practice, these models hold the promise of revolutionizing respiratory care, enabling more comprehensive and accessible assessments of lung function, and ultimately improving patient outcomes. Overall, the models demonstrate robust performance across lung volume measurements, underscoring their potential utility in clinical practice for accurate diagnosis and prognosis of respiratory conditions in locations where access to body plethysmography or other lung volume measurement modalities is challenging..