In an investigation of respiratory health among older adults across India, researchers evaluating the first nationally representative survey with objective lung function testing found that machine learning algorithms provide only modest predictive power when identifying undiagnosed airflow obstruction from routine health measures. The study, led by Rupeshkumar Deshmukh and Sudhanshu Mahajan from Bharati Vidyapeeth Medical College in Pune and published in Frontiers in Artificial Intelligence, analyzed data from the Longitudinal Ageing Study in India Wave 1. Among 35,051 adults aged forty-five and older with complete clinical measurements and usable spirometry, 12.0 percent exhibited spirometry-defined airflow obstruction under the Global Initiative for Chronic Obstructive Lung Disease fixed-ratio criterion. Strikingly, 89.1 percent of these individuals—totaling 3,751 participants—had never received a clinical diagnosis of chronic lung disease. However, despite benchmarking ten advanced statistical and machine learning approaches, model discrimination remained modest, demonstrating that non-invasive patient characteristics alone cannot reliably replace diagnostic spirometry in clinical practice.
Chronic obstructive pulmonary disease represents a leading cause of mortality and disability in India, yet most cases remain unrecognized due to sparse diagnostic infrastructure and reliance on self-reporting. Unlike high-income settings where tobacco smoking dominates airway disease, India faces substantial respiratory morbidity driven by non-smoking exposures, including indoor biomass fuel combustion, occupational dust, and ambient particulate pollution. The Longitudinal Ageing Study in India provides the first population-based framework pairing structured interviews with standardized, pre-bronchodilator handheld ultrasonic spirometry, physical anthropometry, and dried-blood-spot assays. The authors emphasize that because post-bronchodilator testing was not administered during field interviews, an observed pre-bronchodilator ratio of forced expiratory volume in one second to forced vital capacity below 0.70 defines an obstructive spirometric pattern rather than an established clinical diagnosis of chronic obstructive pulmonary disease. Individuals exhibiting this pattern may have asthma, post-tuberculosis sequelae, or transient bronchospasm, underscoring the vital distinction between population-level spirometric impairment and verified clinical disease.
To establish an uncompromised analytic cohort, investigators performed a participant derivation using unique identifiers from the public survey files. Starting from 73,396 individuals across thirty-six states and union territories, age eligibility of forty-five years and above was applied first, excluding 6,790 respondents who were younger spouses. Among 66,606 age-eligible individuals, 60,643 attempted spirometry, while 5,963 could not complete testing due to contraindications, refusals, or incomplete protocols. Within those attempting testing, 12,549 were coded as not usable and 10,340 lacked quality evaluations, leaving 37,754 participants with usable examinations. Further excluding 2,703 individuals with missing predictor data yielded the final analytic sample of 35,051 individuals. The excluded participants did not differ meaningfully in age or obstruction prevalence from the final sample, though urban individuals exhibited slightly higher laboratory missingness. In geographic coding, the investigators maintained Jammu and Kashmir as a single undivided category, reflecting the pre-2019 administrative structure preserved in the public release.
The predictive framework incorporated twenty-six candidate predictors verified directly against survey instruments, intentionally omitting spirometry-derived parameters to simulate realistic primary-care screening. Candidate variables encompassed demographic factors such as age, sex, education, religion, caste, and marital status; socioeconomic variables including urban or rural residence, living arrangements, wealth quintile, and state; lifestyle factors including tobacco and alcohol use alongside vigorous physical activity; health insurance coverage; and clinical measures including blood pressure, pulse rate, body mass index, waist and hip circumferences, dried-blood-spot biomarkers consisting of haemoglobin, glycated haemoglobin, and C-reactive protein, a prior tuberculosis flag, and a comorbidity index. Chronic lung disease was strictly excluded from comorbidity counts to prevent circular leakage. Baseline disparities distinguished obstructed from non-obstructed participants: individuals with obstruction were older, averaging 61.47 years compared to 57.38 years, had lower average body mass index, at 21.99 compared to 23.13 kilograms per square meter, and were substantially more likely to report tobacco use, at 15.3 percent versus 9.9 percent.
Geographic analysis revealed marked heterogeneity in crude obstruction prevalence across India, spanning from 1.8 percent in Meghalaya to 25.1 percent in Himachal Pradesh and 24.1 percent in Jammu and Kashmir, representing an approximate fourteen-fold regional variation. Nationally, while the crude prevalence was 12.0 percent, incorporating survey-specific sampling weights via primary sampling unit cluster bootstrapping yielded a national population estimate of 12.6 percent, with a 95 percent confidence interval between 11.8 and 13.5 percent. The authors cautioned against interpreting individual state differences as definitive evidence of local environmental or behavioral hazards, such as biomass burning or high altitude, because per-state sample sizes were modest, confidence intervals were broad, and individual-level exposure histories could not be disentangled within cross-sectional data. Furthermore, sparse sampling unit counts precluded design-based weighting at the state level. Consequently, the observed geographic variations serve as descriptive baselines for future research rather than validated metrics for direct health resource reallocation.
To benchmark predictive capability, the development sample of 32,028 participants interviewed in 2017 and 2018 was split into an eighty percent training set of 25,622 individuals and a twenty percent internal test set of 6,406 individuals, stratified by outcome prevalence. Ten distinct modeling methodologies were examined, including logistic regression, random forest, AdaBoost, k-nearest neighbours, XGBoost, LightGBM, CatBoost, a deep neural network, a soft-voting ensemble, and a stacked hybrid meta-learner combining cross-validated out-of-fold predictions. Classification thresholds were optimized strictly on training data using Youden index criteria. On the internal test set, discrimination across top-performing algorithms remained modest and tightly clustered: the stacked model achieved an area under the curve of 0.686, followed by CatBoost at 0.685, XGBoost at 0.683, and standard logistic regression at 0.672. A flexible logistic regression model incorporating natural spline terms achieved an area under the curve of 0.683. Paired bootstrap comparisons revealed no statistically significant advantage for complex tree ensembles or stacking over well-specified regression.
The investigation evaluated model robustness through temporal holdout validation, reserving 3,023 participants interviewed between 2019 and 2021 who were unseen during model development. On this temporal cohort, where observed obstruction prevalence declined to 7.8 percent, random forest achieved the highest discrimination with an area under the curve of 0.674, compared to 0.665 for the stacked ensemble and 0.662 for CatBoost, reversing the internal ranking. However, pairwise bootstrap differences demonstrated that performance between leading models remained statistically indistinguishable. The authors noted that 665 validation participants resided in Sikkim, a state interviewed entirely during 2020 and 2021 that was absent from development data. When Sikkim was excluded in sensitivity analysis, CatBoost and the stacked model achieved identical areas under the curve of 0.673, outperforming random forest at 0.665. This shift highlighted how temporal validation rankings were partially confounded by an unobserved geographic category rather than temporal drift alone, underscoring the necessity of handling geographic variables carefully in regional health models.
Model explainability using SHAP values derived from CatBoost identified age, body mass index, state of residence, tobacco use, and sex as the five most influential predictors. The authors explicitly warned that these values reflect model attribution within observational data rather than direct causal effects, especially regarding nominal codes for state and caste. Evaluating model calibration proved equally revealing: training models with class-imbalance weighting produced severe probability distortion without improving discrimination. Weighted CatBoost predicted an average risk of 42.9 percent on the internal test set where actual prevalence was 12.4 percent, whereas unweighted CatBoost predicted 12.3 percent with excellent calibration and a Brier score of 0.103. Absolute probability calibration worsened on temporal validation due to lower baseline prevalence. Furthermore, nested resampling across five outer folds grouped by sampling unit yielded a pooled out-of-fold area under the curve of 0.687, matching the inner selection performance and confirming that model selection and threshold derivation introduced virtually zero statistical optimism.
Seeking an accessible triage tool, the authors formulated a five-item bedside score assigning integer points: three points for age sixty-five or older, two for male sex, two for body mass index below 21 kilograms per square meter, two for tobacco use, and one for waist circumference of ninety centimeters or greater. At the cross-validated threshold of five points or more, internal test sensitivity was 54.0 percent and specificity was 66.0 percent, with predictive values of 18.4 percent positive and 91.0 percent negative, leaving 5.7 cases missed per one hundred screened. In temporal validation, sensitivity fell to 45.5 percent, missing 4.2 cases per one hundred screened. Likelihood ratios hovered close to unity—positive ratios of 1.59 internally and 1.51 temporally, and negative ratios of 0.70 and 0.78—demonstrating negligible shifts in post-test probability. Lowering the cut-point to four points increased sensitivity to 68.9 percent but referred over half the population. Decision curve analysis indicated that risk-based prioritization offered negligible net benefit over universal referral within clinically realistic thresholds.
Crucially, subgroup analysis revealed profound equity limitations in the bedside score, which achieved 73.5 percent sensitivity among men but only 21.1 percent among women. Because the score assigns no points for female sex and relies heavily on smoking and older age, women under sixty-five who never used tobacco can score at most three points, making referral structurally impossible despite representing nearly twenty percent of all obstructed cases. When evaluating an alternative threshold based on the Global Lung Function Initiative 2012 lower limit of normal, obstruction prevalence rose to 14.1 percent, but model discrimination plummeted to approximately 0.61 because age and sex no longer provided diagnostic separation. The researchers concluded that neither complex machine learning nor simple clinical scoring can reliably rule out airflow obstruction or substitute for objective spirometric evaluation. While simple scores might modestly assist in scheduling spirometry in resource-limited facilities, prospective trials must evaluate whether symptom-based referral pathways incorporating cough, breathlessness, and domestic exposures offer a more equitable, clinically sound approach.
Subject of Research: Machine learning evaluation and bedside clinical scoring for predicting spirometry-defined airflow obstruction among older adults in India.
Article Title: Machine learning prediction of spirometry-defined airflow obstruction in older Indian adults: a multimodel benchmark, simplified bedside score, and temporal holdout validation
Article References: Deshmukh, R., & Mahajan, S. (2026). Machine learning prediction of spirometry-defined airflow obstruction in older Indian adults: a multimodel benchmark, simplified bedside score, and temporal holdout validation. Frontiers in Artificial Intelligence, 9, Article 1986141. https://doi.org/10.3389/frai.2026.1986141
Image Credits: AI Generated
DOI: 10.3389/frai.2026.1986141
Keywords: airflow obstruction, spirometry, Longitudinal Ageing Study in India, machine learning benchmark, bedside clinical score, chronic obstructive pulmonary disease, CatBoost, stacked generalisation, temporal validation, underdiagnosis, fixed-ratio criterion, lower limit of normal, decision curve analysis, calibration intercept, TRIPOD-AI guidelines, health screening disparities, respiratory epidemiology
News Source: Teresa Odom. (November 12, 2026). Machine Learning Modestly Predicts Airflow Obstruction in Older Indian Adults. Scienmag.



