Pulmonary function tests (PFTs) are a series of noninvasive procedures that assess the overall function and health of a patient’s lungs. A complete PFT can measure many variables reflecting different aspects of the patient’s ventilatory and gas exchange function. Deviation from population-based normative values is fundamental to the diagnosis, management, and prognosis of a variety of respiratory disorders. Several physiologic and pathologic patterns are generated by combinations of lung function variables from PFTs. In response to this latter concern, the concept of “lung age” derived from PFTs was developed as a means of “summarizing” the patient’s lung function and making it more intuitive to patients [1-7]. Because it is derived from multiple measures from the PFT, lung age may be more reflective of the patient’s overall pulmonary physiology. Scientifically, lung age has shown clinical importance in a randomized controlled trial by increasing smoking cessation [8]. The current methodology to calculate lung age depends on multiple linear regression methods and uses discrete variables obtained from the spirometry and sometimes additional clinical parameters [3,4].

Lung age is a single summary variable useful in patient education, research, and clinical practice [9-13]. For example, in chronic obstructive pulmonary disease (COPD), the forced expiratory volume in 1 second (FEV1) assesses the severity of obstruction, while in idiopathic pulmonary fibrosis (IPF), the forced vital capacity (FVC) reflects lung capacity and guides prognosis and management [10-13]. Although FEV1 and FVC are critical variables in obstructive or restrictive disorders, other variables, such as the residual volume (RV) or the diffusing capacity, can provide additional information on the patient’s respiratory function. It is also not uncommon to have mixed disorders present in the same patient, for example, combined pulmonary fibrosis with emphysema and IPF complicated by pulmonary arterial hypertension [14,15]. While convenient, the use of individual variables may therefore not communicate the breadth of the patient’s lung pathology, and importantly, these variables are conceptually less meaningful to patients who often lack the requisite medical literacy to understand respiratory pathophysiology.

PFTs reveal notable differences between biological sexes due to anatomical and physiological variations, impacting respiratory health. Men typically exhibit larger lung volumes and higher absolute values for FVC and FEV1 even when adjusted for age and height, while women often have smaller airway diameters and reduced expiratory flow rates [16-18]. A population study analyzed 3250 patients and found that sex differences in breathlessness are explained by absolute FEV1 values [16]. Given these differences, machine learning techniques are well-suited to predict patient sex based on pulmonary function test data.

In this investigation, we aimed to create an accurate lung age and sex prediction tool by incorporating all key physiological measures from the complete PFT and applying machine learning methods. We report here the development and validation of a machine learning–assisted lung age and sex prediction tool.

Our study offers an innovative perspective on leveraging machine learning to derive a single, intuitive index—lung age—from PFT data. In doing so, we extend the concept of lung age, which has historically relied on linear regression and a limited number of spirometric variables such as FEV₁ and FVC [1-5,7,8,22,23]. Although traditional lung age estimates have informed clinical practice, particularly in motivating smoking cessation, their scope may be inadequate for detecting mixed respiratory disorders or capturing the complexity of progressive diseases, such as COPD and IPF [1,4,8,12,15,19,22,23]. By integrating additional parameters, including lung volumes and diffusing capacity, our models encapsulate broader physiologic domains, thereby providing clinicians and patients with a more complete snapshot of pulmonary function.

A distinguishing innovation in our methodology lies in the incorporation of time-series features from expiratory flow-volume curves, an approach that harnesses the temporal complexity of airflow mechanics [24]. Traditional linear models often reduce such curves to a few key values (eg, peak expiratory flow or FEV₁/FVC ratio), potentially discarding clinically relevant information hidden within continuous signals [19,25]. The use of gradient-boosted machines enabled the model to use these subtle interactions among multiple respiratory variables, resulting in significantly enhanced predictive accuracy for both age and sex classifications. Interestingly, model 3—which relied solely on tabular PFT data—offers a near-equivalent performance while being more amenable to seamless clinical adoption, such as an online calculator for diverse health care environments. In contrast, the most advanced model (model 4), which incorporated time-series data, exhibited marginally superior accuracy and may be best suited for specialized centers capable of handling more intensive data processing. These findings encourage further research to determine whether the incremental performance gains from time-series data justify the greater computational demand, especially in patient populations with established respiratory dysfunction.

Beyond its use for age estimation, these models demonstrated robust performance in sex classification, underscoring the fundamental physiological differences that exist between male and female lungs [26-28]. This high fidelity not only addresses potential gaps in demographic data encountered in large-scale epidemiological surveys but could also influence more nuanced investigations into sex-specific risks and disease trajectories [29-32]. Additionally, the extension of sex classification models to transgender populations may offer an innovative lens on how hormonal therapies or varying physiological baselines interact with lung function over time, but these questions require meticulous data collection regarding gender identity and transitions. From a clinical use perspective, precise sex classification can refine predictive models, ensuring that normative ranges and targeted interventions align more accurately with individual patient needs.

This work highlights the importance of model interpretability, which is often overlooked in the rapid adoption of machine learning approaches in health care. By leveraging SHAP and partial dependence plots, we were able to offer clinicians clear insights into how specific features (eg, RV, functional residual capacity, anthropometric variables) drive each individual prediction [33,34]. This transparency is pivotal for patient-centered communication, as it enables clinicians to contextualize changes in lung age or sex-based predictions within underlying physiologic processes and to discuss tailored management strategies. By identifying which variables exert the greatest influence on a patient’s estimated lung age, for instance, a provider might more effectively counsel on lifestyle modifications, targeted therapies, or disease-monitoring intervals. For age predictions, SHAP analysis showed that RV as a percentage of TLC, FEV₁, and alveolar volume were among the most influential features. For sex predictions, the SHAP plots showed distinct patterns in how features, such as peak expiratory flow and height, contributed to classification outcomes for male participants versus female participants. Importantly, anthropometric ablation analyses showed only modest reductions in performance after the removal of height and weight, suggesting that body size contributes to prediction but does not fully account for model performance.

Lung volumes, notably the RV and the RV/TLC ratio, increase with advancing age and may reduce the vital capacity [35]. This age-related trend aligns with the physiological principle that dynamic small airway compression caused by loss of alveolar elasticity due to aging traps gas in the lungs and elevates RV [36]. In COPD, resting pulmonary hyperinflation (RV/TLC ≥40%) correlates with older age [37]. If the increase in RV surpasses functional residual capacity, more peripheral small airways may not be ventilated, subsequently failing to participate in gas exchange and affecting the regional ventilation-perfusion ratio, which may cause the diffusion capacity of the lung for carbon monoxide to also decline with age [35]. Cross-sectional studies may overestimate or underestimate these age-related changes because of cohort and period effects, including improved measurement techniques and shifting environmental exposures. Individuals with wheezing symptoms or lower FEV₁ often exhibit higher RV/TLC, underscoring the combined role of airflow obstruction and hyperinflation [35]. Furthermore, a negative relationship between RV/TLC and diffusing capacity adjusted for lung volume (diffusion capacity of the lung for carbon monoxide/alveolar volume) points to alveolar enlargement or membrane alterations that hinder gas exchange [35]. Collectively, these findings emphasize how structural, mechanical, and biochemical factors converge to shape lung aging.

The healthy cohort was skewed toward later adulthood, and age-stratified analyses showed that model error was not uniform across the age spectrum. Both models 3 and 4 tended to overestimate lung age in younger adults and underestimate lung age in older adults, with the smallest errors near the cohort’s median age. This pattern is consistent with regression to the mean in a training distribution concentrated in middle and later adulthood. Accordingly, lung age estimates from the current model should be interpreted most cautiously at the youngest and oldest ages. Future efforts to mitigate this pattern may include age-balanced sampling during model development, expansion of healthy participants at the age extremes, post hoc recalibration of predicted lung age against chronological age, and reliability flagging for underrepresented age ranges.

The sex prediction models also demonstrated high accuracy, with model 4 achieving an AUC of 0.981, indicating robust discrimination between male and female pulmonary physiology. This performance may reflect known sex-related differences in lung size, airway diameter, and volume distribution. Key contributing features in sex prediction included the FEV1/FVC ratio, RV, and functional residual capacity, with demographic factors, such as height and weight, also influencing model predictions. Notably, peak expiratory flow and height showed distinct contributions to male and female classifications, aligning with known sex-based anatomical and physiological variations [17,18,38]. It would be interesting to apply the sex model within the transgender population.

Nonetheless, our findings have certain limitations. The retrospective study design, coupled with the inclusion of patients undergoing PFTs for clinical indications ranging from presurgical evaluation to pre–bone marrow transplant assessment to dyspnea without respiratory disease, may limit broader generalizability. In addition, because the cohort was derived from patients undergoing clinically indicated testing rather than from a population-based healthy sample, some selection bias may remain despite the rigorous healthy cohort definition. The cohort was also predominantly Caucasian, which may limit generalizability to more diverse populations. Moreover, although lung age incorporates data from multiple physiologic dimensions, real-world deployment hinges on the availability of advanced PFT equipment, consistent procedural quality, and robust data integration infrastructures, factors that might challenge adoption in resource-limited settings [39]. Ethical and data governance considerations also arise when implementing machine learning models at scale, given the sensitive nature of electronic health record systems and potential biases introduced by incomplete or inaccurate patient information [40,41].

Lung age models have the potential to change how pulmonary health is assessed, communicated, and monitored. By simplifying pulmonary function test data into a single, intuitive metric, lung age may provide a tool that enhances our understanding of disease severity, improves risk communication, and motivates behavioral change, particularly in high-risk populations such as smokers. These models may also serve as objective biomarkers for early disease detection, longitudinal monitoring, and therapeutic response, enabling clinicians to track subtle physiologic changes over time that may not be readily apparent through traditional spirometric thresholds alone. Future research could expand on these results in several ways. Prospective, longitudinal studies might elucidate how lung age evolves longitudinally over multiple PFTs, particularly in conditions marked by gradual pulmonary decline, such as IPF or advanced COPD [42]. In such contexts, lung age might serve as a surrogate end point for clinical trials, offering a more personalized measure of disease progression and treatment response than single spirometric measures alone [43]. Biomarker or imaging integration represents another promising frontier; coupling lung age with genomic, proteomic, or radiomic data could uncover pathophysiologic pathways of accelerated lung aging or detect preclinical forms of lung disease [44]. Finally, as telehealth platforms and wearable respiratory sensors proliferate, there may be opportunities to develop simplified lung age offshoots—perhaps relying on robust but easily recorded flow-volume data—that can be administered remotely, expanding access to preventive care and early disease detection [45]. There is also potential in the longevity medicine field to create physiologic age scores with the help of models, such as lung age.

In conclusion, this study advances the notion that machine learning can transform large volumes of PFT data into a cohesive and clinically actionable index of respiratory health. The novelty of time-series analysis, combined with interpretability frameworks, promises not only more accurate estimations of lung age but also a deeper understanding of sex-related physiological differences. Over the long term, we postulate that lung age, particularly in its more advanced configurations, could emerge as a powerful surrogate measure for disease severity and prognosis, further refining how clinicians and patients perceive and manage respiratory health. As we continue to validate lung age and sex in broader populations and more diverse clinical contexts, the prospect of integrating machine learning–driven insights into routine pulmonary care draws closer to reality, potentially revolutionizing how we monitor and intervene in chronic lung conditions.