Machine learning prediction of spirometry-defined airflow obstruction in older Indian adults: a multimodel benchmark, simplified bedside score, and temporal holdout validation

Background Chronic obstructive pulmonary disease (COPD) is substantially underdiagnosed in India. The Longitudinal Ageing Study in India (LASI) Wave 1 is the first nationally representative survey to administer objective, pre-bronchodilator spirometry to older Indian adults. We examined how well routinely collected, non-spirometric information identifies spirometry-defined airflow obstruction, and whether a simple bedside score could help prioritise referral for spirometry. Methods We analysed 37,754 LASI Wave 1 respondents aged ≥45 years with usable spirometry, of whom 35,051 had complete data on 26 predictors. Airflow obstruction was defined by the GOLD fixed-ratio criterion (pre-bronchodilator FEV1/FVC < 0.70); because post-bronchodilator testing was not performed, the outcome is spirometry-defined airflow obstruction and not a clinical diagnosis of COPD. Ten prediction approaches, including a stacked ensemble, were developed on 80% of the 2017–18 interviews and tested on the remaining 20%, with classification thresholds chosen by 10-fold cross-validation on the training data only. Interviews from 2019 to 2021 ( n = 3,023) formed a temporal holdout sample, one state of which (Sikkim) was absent from development. We report bootstrap confidence intervals, paired AUC differences, calibration intercept and slope, Brier scores, and net benefit, and a PSU-cluster bootstrap for survey-weighted prevalence. The complete modelling procedure was also evaluated by nested resampling grouped by sampling unit, and a lower-limit-of-normal outcome was analysed separately. The principal analysis was repeated in examinations flagged acceptable, and the lower limits of normal supplied in the data were examined in healthy never-smokers. Sensitivity analyses covered a lower-limit-of-normal outcome, class-imbalance correction, missing data, inverse-probability weighting, laboratory-free models, and a logistic model with spline terms. Reporting followed TRIPOD+AI. Results Airflow obstruction was present in 4,212 of 35,051 participants (12.0%; survey-weighted prevalence 12.6%, 95% CI 11.8–13.5%); only 10.9% of these participants had ever reported a chronic lung disease diagnosis. Discrimination was modest: on the internal test set the AUC was 0.686 for the stacked model and 0.685 for CatBoost (paired difference 0.001, 95% CI −0.002 to 0.004), and on temporal validation 0.674 for random forest, 0.665 for the stacked model, and 0.662 for CatBoost, with no statistically supported differences between the leading models. A logistic regression with spline terms performed similarly (internal AUC 0.683). Obstruction prevalence ranged from 1.8 to 25.1% across states. Age, BMI, state, tobacco use, and sex were the leading predictors. A five-item bedside score (age ≥65 years, male sex, BMI < 21 kg/m 2 , ever tobacco use, waist ≥90 cm) had an AUC of 0.633 internally and 0.621 on temporal validation; at the cross-validated threshold of ≥5 points, sensitivity was 54.0 and 45.5%, 5.7 and 4.2 cases per 100 people screened were missed, and the positive predictive value was 18.4 and 11.3%. Sensitivity was 73.5% in men and 21.1% in women. At the cross-validated thresholds, accuracy on the internal test set was 0.637 (95% CI 0.626–0.649) for the stacked model and 0.657 (0.646–0.667) for CatBoost, with sensitivity of about 60% and specificity of about 65%; accuracy was 0.749 (0.731–0.762) for CatBoost on temporal validation, and it reflected the operating threshold more than the ability to rank participants. Class weighting did not improve discrimination but inflated predicted risks about threefold, and decision curve analysis showed no consistent advantage of risk-based referral over referring everyone for spirometry at threshold probabilities of 5 to 10%. Nested resampling of the whole procedure gave a pooled AUC of 0.687 (95% CI 0.677 to 0.697), with no optimism from model selection. With a lower-limit-of-normal definition, prevalence was 14.1%, the kappa with the fixed ratio was 0.81, and the AUC fell to 0.61 (age and sex alone: 0.52); the bedside score had no useful discrimination for that outcome. Restricting the analysis to the 28,805 participants whose examinations were flagged acceptable gave similar AUCs (internal test 0.679 to 0.689; temporal validation 0.646 to 0.671). Among healthy never-smokers, 11.5% fell below the lower limit of normal supplied in the file, against 5% expected. Conclusion Routinely collected non-invasive information discriminates spirometry-defined airflow obstruction only modestly, and complex machine-learning models showed no demonstrable advantage over a well-specified logistic model. The bedside score misses roughly half of the people with obstruction and is least sensitive in women and younger never-smokers; it should not be used to exclude obstruction or to replace spirometry or clinical judgement. At most, it could serve as a preliminary aid for prioritising spirometry where capacity is limited, pending prospective evaluation against a symptom-based pathway. Findings apply to the fixed-ratio definition and not to a lower-limit-of-normal definition, and to a single survey wave.