Abstract
Background and objectives
Native and indigenous communities have a high burden of chronic kidney disease (CKD). The burden of this condition amongst the indigenous populations of India is poorly documented. We estimated the prevalence and determinants of impaired kidney function among indigenous adults in Odisha, using large scale sample survey data, employing machine learning models.
Methods
We analysed data from 19,703 adults across 53 indigenous communities enrolled in the Odisha Tribal Family Health Survey (OTFHS, 2022–23). Data were collected with serum creatinine, from which estimated glomerular filtration rate (eGFR) was derived using chronic kidney disease epidemiology collaboration equation (CKD-EPI). Impaired kidney function (IKF) was defined as eGFR <90 mL/min/1.73m2 and classified into mild, moderate, severe, and failure categories. Deep neural network (DNN) and eXtreme gradient boosting (XGBoost) models were developed for early prediction for IKF.
Results
Overall, 24.8% [95% confidence interval (CI): 23.8–25.9] of participants had IKF: 16.6% mild, 6.4% moderate, 1.5% severe, and 0.3% kidney failure. The prevalence increased with age (5.6% in 18–30 y vs. 79.4% in >70 y) and was higher among men, those with lower education, tobacco or alcohol use, and individuals with hypertension, diabetes, or underweight. The DNN model showed good predictive performance (accuracy 0.78; recall 0.69), with improved recall (0.78) at a threshold of 0.4. Decision curve analysis suggested a potential benefit of machine learning (ML) based targeted screening. Shapley Additive explanations (SHAP) analysis identified age, hypertension, diabetes, tobacco use, alcohol use, and low dietary diversity as key predictors.
Interpretations and conclusions
Nearly one in four adults from indigenous communities in Odisha had IKF, highlighting a substantial and inequitable kidney disease burden. Machine learning models using routine data offer a feasible strategy for early detection and targeted interventions in resource-limited indigenous settings.
Keywords: Chronic kidney disease, Impaired kidney function, Indigenous health, Machine learning, OTFHS
Indigenous populations in India experience a disproportionate burden of disease, with higher rates of morbidity, mortality, and reduced life expectancy.1,2 Common health challenges among India’s indigenous populations include the so-called ‘quadruple burden’: infectious and communicable diseases (e.g., tuberculosis, malaria), undernutrition (notably among women and children), and growing rates of non-communicable diseases (NCDs), substance use and mental health concerns.3 National surveys and studies show an increasing prevalence of NCDs, including chronic kidney disease (CKD), which has emerged as a pressing public health concern.4-6 In addition to the traditional risk factors (diabetes, hypertension), CKD risk is also shaped by contextual factors such as geography (altitude, water quality), environmental exposures, occupational practices (especially in agriculture and mining), socio-economic deprivation, educational attainment, and dietary patterns.6 These risk factors disproportionate affect rural and indigenous populations.7,8
The true burden of CKD in indigenous populations is underestimated due to low awareness, limited diagnostic access, and often silent progression until late stages, especially in rural and indigenous settings.9 Given these multi-layered risks, there is a crucial need to understand the disease burden so that proactive, context-sensitive approaches can be developed for timely identification and management. Predictive modelling using machine learning (ML) offers a robust approach for leveraging routine sociodemographic and clinical survey data to estimate individual and population-level disease risk.10,11 In public health, ML techniques are increasingly used to predict disease risk, inform early detection, improve screening efficiency, and tailor preventive interventions.12 For a range of chronic diseases such as diabetes, cardiovascular disease, and chronic obstructive pulmonary disease, risk modelling has been instrumental in stratifying populations and optimising resource allocation for disease control.13
In the context of indigenous health, predictive modelling holds special promise for bridging gaps in early disease detection, especially for conditions like impaired kidney function, where progression is often silent, insidious, and frequently irreversible without timely intervention. By capturing complex interrelations among demographic, environmental, and lifestyle risk factors, ML-powered models can identify high-risk individuals and prompt targeted outreach, potentially transforming population health for India’s indigenous people. This study aimed to estimate the prevalence and determinants of impaired kidney function among adults from indigenous communities in Odisha and to develop and validate machine learning models for early prediction using large-scale sample survey data. We hypothesised that ML would allow us to utilise routinely collected data to estimate the burden of kidney injury and understand the interaction of structural inequalities, environmental exposures, and social determinants of health that create unique pathways to kidney injury.
Methods
This survey based study was undertaken by the ICMR-National Institute of Health Research, Bhubaneswar, Odisha, India, after obtaining the ethical clearance from the Institutional Ethics committee of the institute.
Study setting
The study was carried out in 14 districts designated as ‘tribal’ due to their high concentration of Indigenous communities in the State of Odisha, in eastern India.
Study design and participants
We used data from the Odisha Tribal Family Health Survey (OTFHS), undertaken between July 2022 and July 2023. The detailed methodology of the survey and summary findings are published elsewhere.14,15 Briefly, clusters (villages/hamlets/urban blocks) were chosen using multistage cluster sampling and probability proportional to size design. The primary sampling units were clusters, defined as villages, notified hamlets, or urban census blocks. These clusters were selected using probability proportional to size (PPS) sampling based on their population size. Households served as the secondary sampling units and were chosen through systematic random sampling. In each selected household, all permanent indigenous residents were enrolled. Separate sample size estimates for each tribe ensured subgroup validity. The survey covered 389 clusters, 10,090 households, and 30,292 individuals from 53 indigenous communities. Pregnant women and those with active malignancy, or those on immunosuppressive therapy were excluded.
Tools and definitions
Trained staff used structured questionnaires for sociodemographic, behavioural, and medical history data. Anthropometric measures (height, weight, body mass index, waist/hip ratio) and blood pressure (average of two seated readings) were taken using standard protocols. Hypertension and diabetes were defined by clinical thresholds or current medication.16,17 BMI categories followed the WHO Asia-Pacific standards.18 Socioeconomic status was determined via a principal component analysis-derived wealth index.19 Dietary diversity score (DDS) was calculated using standard methodology, and individuals were categorised into low (≤3), medium (>3–5), and high (>5) dietary diversity groups.20
Adults (>18 y) provided blood samples. Serum creatinine and glucose testing were done in a central laboratory using validated methodology on an EM 360 analyser, adhering to internal and external quality protocols. Estimated glomerular filtration rate (eGFR) was calculated by the chronic kidney disease epidemiology collaboration equation (CKD-EPI) equation. Impaired kidney function (IKF) was defined as mild (eGFR 60-89 mL/min/1.73m2), moderate (30-59 mL/min/1.73m2), severe (15-29 mL/min/1.73m2), or failure (<15 mL/min/1.73m2). Only complete biomarker datasets were analysed.
Classification models
Two machine learning models, deep neural network (DNN) and eXtreme gradient boosting (XGBoost), were developed to predict IKF using sociodemographic, anthropometric, and clinical variables (excluding biomarkers used in calculating eGFR). The dataset was split (70% training, 30% testing); 20% of the training subset served for internal validation. Feature selection combined domain expertise and data exploration. Model development followed best practices for preprocessing, handling class imbalance, and hyperparameter tuning.
Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the curve (AUC) metrics. We explored various threshold probabilities to balance sensitivity and specificity, reporting results at thresholds of 0.5, 0.45, 0.4, and 0.3 confusion matrices. Receiver operating characteristics (ROC) curves were generated accordingly.
Decision curve and model interpretation
Decision curve analysis evaluated the net benefit of prediction-based vs. universal or no screening across probabilities.21 Feature importance was analysed using Shapley Additive explanations (SHAP) values, identifying key predictors and high-risk phenotypes. Based on performance and interpretability, the DNN at threshold 0.4 was optimal. Spatial patterns were mapped with QGIS v3.34.
Results
Sample characteristics
A total of 19,703 adults from 53 indigenous communities across 14 districts of Odisha were included in the analysis. The mean age was 41.7 y (SD 15.9), and 57.9% were female. Tobacco use was common, with more than half of participants reporting current use, while about one-quarter reported alcohol consumption. Nearly one-third had no formal education, and agriculture was the predominant occupation.
Prevalence of impaired kidney function
Overall, 24.8% [95% confidence interval (CI): 23.8–25.9%] of participants exhibited IKF, with most cases classified as mild or moderate ( Table I). A strong age gradient was observed, with prevalence increasing markedly in older age groups. Men had a higher prevalence of IKF compared with women, and scheduled tribes showed slightly higher prevalence than particularly vulnerable tribal groups (PVTGs). Socioeconomic differences were also evident. IKF was more common among individuals with lower educational attainment and those belonging to lower wealth quintiles. Similarly, higher prevalence was observed among individuals reporting tobacco and alcohol use ( Table I).
Table I.
Distribution of impaired kidney function across different characteristics of the indigenous people of Odisha, 2022-23 (N=19703)
| Characteristics |
n (%) | Kidney function categories (%, 95% CI) |
|||||
|---|---|---|---|---|---|---|---|
| Normal | Mild IKF | Moderate IKF | Severe IKF | Kidney failure | |||
| Sex | Male | 8296 (42.1) | 72.3 (70.5, 74.1) | 18.1 (16.5, 19.8) | 7.3 (6.51, 8.29) | 2.0 (1.60, 2.39) | 0.3 (0.18, 0.42) |
| Female | 11407 (57.9) | 77.5 (76.3, 78.7) | 15.4 (14.4, 16.5) | 5.7 (5.07, 6.40) | 1.1 (0.84, 1.44) | 0.2 (0.14, 0.38) | |
| Age (y) | 18-30 | 5790 (29.4) | 94.4 (93.0, 95.6) | 4.7 (3.58, 6.10) | 0.8 (0.53, 1.26) | 0.1 (0.02, 0.17) | 0.0 (<0.01, <0.01) |
| 31-40 | 4468 (22.7) | 85.0 (83.3, 86.6) | 11.6 (10.2, 13.2) | 2.9 (2.19, 3.74) | 0.3 (0.16, 0.70) | 0.2 (0.05, 0.57) | |
| 41-50 | 3585 (18.2) | 72.8 (70.2, 75.3) | 19.7 (17.6, 22.1) | 5.9 (4.79, 7.26) | 1.3 (0.82, 2.10) | 0.2 (0.09, 0.51) | |
| 51-60 | 2911 (14.8) | 60.3 (57.1, 63.5) | 26.2 (23.2, 29.3) | 10.8 (9.26, 12.6) | 2.6 (1.89, 3.45) | 0.1 (0.06, 0.31) | |
| 61-70 | 1914 (9.7) | 48.5 (44.7, 52.3) | 29.3 (26.0, 32.8) | 16.4 (13.9, 19.2) | 4.9 (3.78, 6.25) | 1.0 (0.58, 1.70) | |
| >70 | 1035 (5.3) | 20.6 (16.9, 25.0) | 45.9 (41.1, 50.8) | 26.0 (21.8, 30.8) | 6.2 (4.24, 8.94) | 1.2 (0.75, 1.95) | |
| Ethnic group | Scheduled tribes | 16047 (81.4) | 74.2 (73.0, 75.4) | 16.8 (15.7, 17.9) | 7.0 (6.38, 7.67) | 1.7 (1.45, 2.04) | 0.3 (0.18, 0.36) |
| PVTG | 3656 (18.6) | 78.9 (77.1, 80.7) | 16.0 (14.5, 17.6) | 4.3 (3.52, 5.18) | 0.6 (0.36, 0.88) | 0.2 (0.09, 0.57) | |
| Wealth quintiles | Lowest | 2954 (15.0) | 78.1 (75.6, 80.4) | 15.1 (13.1, 17.2) | 5.6 (4.53, 6.91) | 1.0 (0.67, 1.50) | 0.3 (0.10, 0.79) |
| Second | 3483 (17.7) | 73.2 (70.8, 75.4) | 17.8 (15.9, 19.8) | 7.0 (5.87, 8.39) | 1.7 (1.11, 2.64) | 0.3 (0.15, 0.49) | |
| Middle | 4044 (20.5) | 72.5 (70.0, 74.9) | 19.1 (16.8, 21.5) | 6.3 (5.35, 7.41) | 1.8 (1.29, 2.44) | 0.4 (0.21, 0.68) | |
| Fourth | 4288 (21.8) | 75.8 (73.5, 78.0) | 16.1 (14.0, 18.3) | 6.7 (5.64, 7.98) | 1.2 (0.84, 1.63) | 0.2 (0.12, 0.46) | |
| Highest | 4934 (25.0) | 76.7 (74.5, 78.8) | 15.1 (13.4, 17.1) | 6.4 (5.22, 7.77) | 1.6 (1.21, 2.12) | 0.1 (0.06, 0.32) | |
| Occupation | Agriculture | 12196 (61.9) | 75.0 (73.6, 76.3) | 17.1 (16.0, 18.3) | 6.2 (5.52, 6.93) | 1.5 (1.17, 1.84) | 0.2 (0.14, 0.38) |
| Labourer | 505 (2.6) | 78.7 (72.8, 83.5) | 11.9 (8.58, 16.2) | 6.9 (4.33, 11.0) | 2.5 (1.25, 5.09) | 0.0 (<0.01, <0.01) | |
| None/homemaker | 5792 (29.4) | 74.2 (72.5, 75.9) | 16.5 (15.1, 17.9) | 7.3 (6.44, 8.34) | 1.6 (1.25, 2.06) | 0.4 (0.25, 0.58) | |
| Retired | 42 (0.2%) | 41.1 (24.1, 60.5) | 23.7 (12.0, 41.4) | 30.8 (15.2, 52.3) | 4.4 (1.31, 13.7) | 0.0 (<0.01, <0.01) | |
| Others | 1168 (5.9) | 79.5 (73.1, 84.7) | 15.7 (10.9, 22.3) | 4.2 (2.55, 6.75) | 0.6 (0.30, 1.06) | 0.0 (<0.01, 0.34) | |
| Diet diversity score | Low dietary diversity (≤3) | 3241 (16.4) | 72.9 (70.0, 75.6) | 18.5 (16.0, 21.3) | 6.7 (5.53, 8.13) | 1.6 (1.02, 2.40) | 0.3 (0.16, 0.49) |
| Medium dietary diversity (3.5–5) | 14045 (71.3) | 74.7 (73.5, 75.9) | 17.0 (15.9, 18.1) | 6.6 (5.98, 7.29) | 1.5 (1.22, 1.77) | 0.2 (0.16, 0.34) | |
| High dietary diversity (>5) | 2417 (12.3) | 80.8 (77.7, 83.6) | 12.3 (9.93, 15.0) | 5.1 (3.95, 6.64) | 1.5 (0.92, 2.42) | 0.3 (0.08, 1.09) | |
| Current tobacco user | Yes | 10106 (51.70) | 69.2 (67.6, 70.9) | 19.6 (18.2, 21.1) | 8.6 (7.73, 9.49) | 2.3 (1.90, 2.77) | 0.3 (0.19, 0.45) |
| Current alcohol user | Yes | 5474 (28.4) | 71.1 (69.2, 72.8) | 18.6 (17.1, 20.2) | 7.9 (7.02, 8.98) | 2.1 (1.63, 2.72) | 0.3 (0.15, 0.50) |
| Education | No education | 10272 (52.4) | 66.7 (65.2, 68.2) | 21.6 (20.2, 23.0) | 9.2 (8.43, 10.1) | 2.1 (1.72, 2.54) | 0.4 (0.24, 0.52) |
| Primary (class 1-5) | 2564 (13.1) | 77.0 (74.0, 79.8) | 15.5 (13.3, 17.9) | 6.1 (4.39, 8.47) | 1.0 (0.60, 1.54) | 0.4 (0.17, 0.86) | |
| Secondary (class 6-10) | 5135 (26.2) | 85.3 (83.6, 86.8) | 10.7 (9.32, 12.2) | 2.9 (2.31, 3.71) | 1.1 (0.75, 1.51) | 0.1 (0.01, 0.23) | |
| Higher secondary and above | 1642 (8.4) | 89.6 (85.4, 92.7) | 8.2 (5.30, 12.5) | 2.0 (1.26, 3.15) | 0.1 (0.06, 0.34) | 0.1 (0.02, 0.36) | |
| Total | 19703 | 75.2 (74.1, 76.2) | 16.6 (15.7, 17.6) | 6.4 (5.93, 7.01) | 1.5 (1.27, 1.75) | 0.3 (0.18, 0.35) | |
PVTG, particularly vulnerable tribal groups; CI, confidence interval
Associations with hypertension, diabetes, and BMI
The prevalence of IKF was substantially higher among participants with hypertension and diabetes ( Table II). Among those with uncontrolled hypertension (raised blood pressure despite on medications), 55.5% had IKF compared with 41.4% in controlled hypertension and 34.7% in those without hypertension. A similar pattern was noted for diabetes: individuals with uncontrolled diabetes had the highest proportion of moderate-to-severe IKF (19.0%), followed by controlled diabetes (11.5%), and non-diabetics (9.9%). BMI showed an association with kidney function: underweight participants had a higher prevalence of moderate or severe IKF (12.9%) compared with those with normal weight (8.6%) or overweight/obesity (6.0%).
Table II.
Relationship of impaired kidney function and diabetes, hypertension and BMI in indigenous people of Odisha
| eGFR Category | Hypertension |
Diabetes |
Body mass index |
No diabetes or hypertension | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Diagnosed |
Undiagnosed | Diagnosed |
Undiagnosed | Underweight | Normal | Overweight | Obesity | ||||
| Under control | Not under control | Under control | Not under control | ||||||||
| Normal | 58.6 (51.2, 66.1) | 44.5 (36.2, 52.7) | 65.3 (63.0, 67.6) | 59.7 (44.2, 73.5) | 55.2 (37.6, 71.6) | 65.3 (60.8, 69.6) | 70.2 (68.1, 72.1) | 77.5 (76.1, 78.8) | 76.7 (72.5, 80.3) | 77.8 (74.0, 81.2) | 80.9 (79.7, 82.1) |
| Mild impaired kidney function | 25.6 (19.3, 31.8) | 35.5 (26.3, 44.6) | 22.7 (20.6, 24.9) | 19.4 (11.8, 30.1) | 33.9 (19.3, 52.4) | 22.0 (18.3, 26.2) | 17.9 (16.3, 19.6) | 15.5 (14.3, 16.8) | 17.3 (13.9, 21.3) | 17.4 (14.2, 21.1) | 13.1 (12.1, 14.2) |
| Moderate impaired kidney function | 13.1 (7.4, 18.9) | 14.2 (9.6, 18.8) | 9.2 (8.02, 10.5) | 17.6 (7.73, 35.1) | 9.1 (4.30, 18.3) | 8.4 (6.71, 10.5) | 9.0 (7.87, 10.3) | 5.7 (5.02, 6.42) | 5.0 (3.60, 6.91) | 4.2 (3.20, 5.61) | 4.9 (4.34, 5.53) |
| Severe impaired kidney function | 1.74 (0.71, 2.7) | 4.9 (2.1, 7.9) | 2.3 (1.82, 2.98) | 1.4 (0.51, 3.62) | 1.6 (0.22, 10.9) | 3.8 (2.40, 5.88) | 2.6 (2.09, 3.15) | 1.1 (0.86, 1.45) | 0.9 (0.38, 2.09) | 0.4 (0.20, 0.91) | 0.9 (0.75, 1.20) |
| Kidney failure | 0.9 (0.07, 1.7) | 0.8 (0.2, 1.3) | 0.5 (0.31, 0.76) | 2.0 (0.63, 6.04) | 0.2 (0.02, 1.20) | 0.5 (0.22, 1.05) | 0.4 (0.24, 0.62) | 0.2 (0.12, 0.36) | 0.1 (0.04, 0.49) | 0.1 (0.03, 0.56) | 0.1 (0.06, 0.24) |
| Total | 507 | 605 | 5366 | 152 | 99 | 1095 | 5957 | 12129 | 1413 | 198 | 11854 |
Values are presented as prevalence (%) with 95% CI
Machine learning model performance
Both DNN and XGBoost models demonstrated moderate predictive performance ( Table III). At the default threshold of 0.5, the DNN achieved an accuracy of 0.78, precision of 0.55, recall of 0.69, and F1-score of 0.61; corresponding values for XGBoost were 0.79, 0.58, 0.62, and 0.60. Lowering the threshold to 0.4 improved recall (0.78 for DNN; 0.70 for XGBoost) at the cost of precision. Confusion matrices across thresholds illustrated the trade-off between false positives and false negatives ( Figs. 1A and B). ROC curves confirmed comparable discriminative ability, with AUCs in the moderate-to-good range ( Fig. 2).
Table III.
Performance evaluation metrics of the DNN and XGboost models for binary classification on the test dataset
| Model | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|
| Default threshold = 0.5 | ||||
| DNN | 0.78 | 0.55 | 0.69 | 0.61 |
| XGboost | 0.79 | 0.58 | 0.62 | 0.60 |
| Threshold = 0.45 | ||||
| DNN | 0.76 | 0.51 | 0.74 | 0.60 |
| XGboost | 0.78 | 0.55 | 0.66 | 0.60 |
| Threshold = 0.4 | ||||
| DNN | 0.74 | 0.48 | 0.78 | 0.60 |
| XGboost | 0.76 | 0.52 | 0.70 | 0.60 |
| Threshold = 0.3 | ||||
| DNN | 0.68 | 0.43 | 0.86 | 0.57 |
| XGboost | 0.72 | 0.46 | 0.77 | 0.58 |
DNN, deep neural network; XGboost, eXtreme gradient boosting
Fig. 1.
Confusion matrices for predictions on the test dataset at different probability thresholds. (A) Deep neural network (DNN) model. (B) Extreme gradient boosting (XGBoost) model. Within each panel, matrices are arranged clockwise from the upper left according to probability thresholds of 0.50, 0.45, 0.40, and 0.30.
Fig. 2.
Receiver Operating Characteristic (ROC) curves, along with the respective area under the curves (AUCs) based on training and testing datasets for the two models.
Decision curve analysis
Decision curve analysis demonstrated that both models provided a clear net benefit compared with ‘test all’ or ‘no test’ strategies across a range of clinically relevant thresholds, with the DNN showing slightly higher utility at lower thresholds ( Fig. 3).
Fig. 3.
Decision curves for the DNN and XGBoost models showing expected net benefits for the strategies: assume all individual will be clinically tested (Test all) and model-based prediction of high-risk individuals who will be clinically tested for IKF—per individual relative to the scenario assuming no individual will be clinically tested (No screening). Unit is the expected benefit associated with 1 individual who would have IKF and would get clinically tested.
Model interpretation and feature importance
SHAP analyses identified age as the major non-modifiable predictor and hypertension, diabetes, BMI, tobacco use, alcohol use, and socioeconomic indicators as the strongest modifiable predictors of IKF. This was consistent across both DNN and XGBoost ( Fig. 3). Age had the largest effect size, with risk increasing sharply beyond 50 y. The models confirm known clinical drivers but also highlight the role of modifiable behavioural and social risks (tobacco, alcohol, low diet diversity, poverty).
Spatial distribution
Mapping demonstrated clustering of IKF in districts with high prevalence of hypertension and diabetes, underscoring shared geographic vulnerabilities and overlapping risk environments (Supplementary Figure).
Supplementary Figure
Discussion
This study provides comprehensive insights into kidney health among Indigenous populations. Our findings indicate a substantial burden of impaired kidney function among adults from indigenous communities in Odisha, with clear demographic, socioeconomic, and behavioural gradients. Applying machine learning methods, we demonstrated the feasibility of risk prediction using routine demographic and clinical survey data, with deep neural networks offering practical thresholds with implications for community screening. Geospatial analyses additionally revealed overlapping hotspots of IKF, hypertension, and diabetes, pointing toward shared genetic, risk factors, and environmental determinants.
Risk factor associations in our study are on expected lines. Strong associations of reduced kidney function with age, hypertension and diabetes are well established.22-24 Our finding of higher IKF among underweight participants extends prior observations from rural India, where both undernutrition and overnutrition coexist as kidney risks.24-26 The role of tobacco use, alcohol consumption, dietary diversity, poverty, and low education also aligns with broader evidence highlighting the importance of social and behavioural determinants of kidney health. Our findings therefore represent the first large-scale epidemiological assessment of kidney impairment burden in indigenous groups in India, and indeed, worldwide, whereas existing studies are small-scale reports primarily focused on identifying CKD of unknown aetiology.27-29 Our use of ML prediction models using survey data is also unique, as most prior CKD risk models rely on clinical cohorts.30,31 By demonstrating moderate predictive performance using routine screening variables, we show that kidney impairment in low-resource indigenous contexts can be detected early, cheaply, and at scale. These findings suggest that while the models are not suited for definitive diagnosis, they perform well as screening tools. The DNN at a threshold of 0.4 provided the best balance of sensitivity and specificity, making it particularly suitable for community-level risk stratification where missing cases are a greater concern than over-referral. This is consistent with broader evidence that ML models, while not substitutes for definitive diagnosis, can serve as efficient, low-cost screening aids.32,33 Decision curve analysis further suggests that ML-guided targeted screening would be more efficient than blanket population screening in indigenous districts where laboratory resources are limited.
Strengths of this study include the large, representative sample, systematic inclusion of multiple tribes across indigenous -dominated districts, and integration of ML tools alongside classical epidemiological analysis. By adding interpretability (decision curves, SHAP values), findings are more policy-relevant and actionable. However, limitations must be acknowledged. As a cross-sectional survey with one-timepoint eGFR measurement, our estimates reflect ‘impaired kidney function,- rather than CKD, which requires repeat testing. This approach may overestimate chronic disease burden by including acute or reversible impairments. While many cited studies here use a similar approach, equivalence with CKD burden cannot be made.27-31 On the other hand, we did not have data on proteinuria since this data was not collected in the survey. This could have led to missing out on cases with early CKD who still did not have eGFR declines. Self-reported tobacco and alcohol use may be subject to misclassification. While predictive models showed feasibility, their moderate accuracy calls for external validation before clinical adoption. Environmental determinants (e.g., water contaminants, heat stress, pesticide exposures) were not measured here but remain likely contributors in indigenous settings.
The high prevalence of IKF identified here underscores the urgent need for early detection and prevention initiatives in indigenous communities. ML tools could be embedded within primary health systems and frontline worker platforms (e.g., tablets used by village health workers for overall NCD screening) to support risk stratification and referral for confirmatory testing. Given the significant overlap between IKF, hypertension, and diabetes hotspots, there is clear scope for geographically prioritised deployment of strengthened, integrated, and intensive programme, particularly focusing on undiagnosed and uncontrolled cases. Complementary prevention strategies aimed at modifiable risk factors such as culturally tailored interventions to reduce tobacco and alcohol use, improve hydration practices in farming labour, and increased dietary diversity need to be implemented, in addition to the use of evidence-guided interventions as suggested in the WHO ‘Best Buys’. Overlapping hotspots suggest that interventions for CKD can be strengthened by integrating them with existing NCD programme and prioritising high-burden districts. Future work includes longitudinal follow up of Odisha Tribal Family Health Survey (OTFHS) participants to assess progression from IKF to established CKD, and to clarify causal pathways.
शोध-संदेश
ओडिशा की आदिवासी आबादी में लगभग हर चार में से एक वयस्क व्यक्ति में किडनी कार्यक्षमता में कमी (impaired kidney function) पाई गई। अध्ययन ने किडनी रोग के जोखिम वाले लोगों की शुरुआती पहचान के लिए मशीन लर्निंग (machine learning) आधारित मॉडल की उपयोगिता दिखाई, जो सीमित संसाधनों वाले क्षेत्रों में समय पर जांच और लक्षित स्वास्थ्य सेवाएं उपलब्ध कराने में सहायक हो सकते हैं।
Acknowledgment
Scheduled Castes and Scheduled Tribes Research and Training Institute (SCSTRTI), the Ministry of Tribal Affairs, and the Government of Odisha for facilitating this study.
Footnotes
How to cite this article: Kshatri JS, Deo V, AK K, Rehman T, Jha V, Pati S. Prevalence, risk-factors and early prediction of impaired kidney function among indigenous adults in Odisha, India: A secondary analysis. Indian J Med Res. 2026;164:621-9. doi: 10.25259/IJMR_491_2026
Author contributions
SP, VJ, VD, JSK: Conceptualisation; JSK, VD, KAK: Data analysis; SP, VJ, JSK: Supervised the study, secured funding; JSK, TR, KAK: Contributed to filed implementation, data collection, project management; JSK, VJ, VD, KAK: Manuscript writing; TR: Assisted with data analysis, manuscript writing. All authors have read and approved the final printed version of the manuscript.
Financial support and sponsorship
The study received funding support from the Scheduled Castes and Scheduled Tribes Research and Training Institute, Government of Odisha (No 1399/SCSTRTI-RES-RES 01/21).
Conflicts of Interest
None.
Use of Artificial Intelligence (AI)-Assisted Technology for manuscript preparation
The authors confirm that there was no use of AI-assisted technology for assisting in the writing of the manuscript and no images were manipulated using AI.
References
- 1.Anderson I, Robson B, Connolly M, Al-Yaman F, Bjertness E, King A, et al. Indigenous and tribal peoples’ health (The Lancet-Lowitja institute global collaboration): A population study. Lancet. 2016;388:131–57. doi: 10.1016/S0140-6736(16)00345-7. [DOI] [PubMed] [Google Scholar]
- 2. World Health Organization. Historic resolution calls for action to improve the health of indigenous peoples. Available from: https://www.who.int/news/item/29-05-2023-historic-resolution-calls-for-action-to-improve-the-health-of-indigenous-peoples, accessed on March 24, 2026.
- 3.Kumar MM, Pathak VK, Ruikar M. Tribal population in India: A public health challenge and road to future. J Family Med Prim Care. 2020;9:508–12. doi: 10.4103/jfmpc.jfmpc_992_19. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. International Institute for Population Sciences (IIPS). Department of Family and Health Welfare, Government of India. National Family Health Survey (NFHS-5). Available from: https://www.nfhsiips.in/nfhsuser/index.php. accessed on March 25, 2026.
- 5.Rekha S, Shirisha P, Muraleedharan VR, Vaidyanathan G, Dash U. Wealth inequalities in nutritional status among the tribal under-5 children in India: A temporal trend analysis using NFHS data of Jharkhand and Odisha states - 2006-21. Dialogues Health. 2023;2:100135. doi: 10.1016/j.dialog.2023.100135. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Kazancioğlu R. Risk factors for chronic kidney disease: An update. Kidney Int Suppl (2011). 2013;3:368–71. doi: 10.1038/kisup.2013.79. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Ghelichi-Ghojogh M, Fararouei M, Seif M, Pakfetrat M. Environmental factors and chronic kidney disease: A case-control study. Sci Rep. 2024;14:26511. doi: 10.1038/s41598-024-72685-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Fitria L, Prihartono NA, Ramdhan DH, Wahyono TYM, Kongtip P, Woskie S. Environmental and occupational risk factors associated with chronic kidney disease of unknown etiology in west Javanese rice farmers, Indonesia. Int J Environ Res Public Health. 2020;17:4521. doi: 10.3390/ijerph17124521. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Francis A, Harhay MN, Ong ACM, Tummalapalli SL, Ortiz A, Fogo AB, et al. Chronic kidney disease and the global public health agenda: An international consensus. Nat Rev Nephrol. 2024;20:473–85. doi: 10.1038/s41581-024-00820-6. [DOI] [PubMed] [Google Scholar]
- 10.Eloranta S, Boman M. Predictive models for clinical decision making: Deep dives in practical machine learning. J Intern Med. 2022;292:278–95. doi: 10.1111/joim.13483. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Manzoor MF. Machine learning for early disease diagnosis: A review of techniques in healthcare applications. Prem J Sci. 2024;5:100043. doi: 10.70389/pjs.100043. [DOI] [Google Scholar]
- 12.Qasrawi R, Issa G, Thwib S, AbuGhoush R, Amro M, Ayyad R, et al. The role of machine learning in infectious disease early detection and prediction in the MENA region: A systematic review. Informatics in Medicine Unlocked. 2025;56:101651. doi: 10.1016/j.imu.2025.101651. [DOI] [Google Scholar]
- 13.Alanazi R. Identification and prediction of chronic diseases using machine learning approach. J Healthc Eng. 2022;2022:2826127. doi: 10.1155/2022/2826127. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Kshatri JS, Mansingh A, Kavitha AK, Bhattacharya H, Bhuyan D, Bhattacharya D, et al. Odisha tribal family health survey: Methods, tools, and protocols for a comprehensive health assessment survey. Front Public Health. 2023;11:1157241. doi: 10.3389/fpubh.2023.1157241. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Kshatri JS, Ak K, Rehman T, Bhattacharya H, Bhuyan D, Mansingh A, et al. A comprehensive assessment of health indicators among tribal populations in Odisha, India (Odisha Tribal Family Health Survey): A community-based, cross-sectional study. Lancet Reg Health Southeast Asia. 2025;38:100611. doi: 10.1016/j.lansea.2025.100611. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. World Health Organization. Hypertension. Available from: https://www.who.int/news-room/fact-sheets/detail/hypertension?utm_source=chatgpt.com, accessed on April 9, 2026.
- 17.Committee ADAPP. ElSayed NA, Aleppo G, Aroda VR, Bannuru RR, Brown FM, et al. 2 Diagnosis and classification of diabetes: Standards of care in diabetes-2024. Diabetes Care. 2024;47:S20–42. doi: 10.2337/dc24-S002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Lim JU, Lee JH, Kim JS, Hwang YI, Kim TH, Lim SY, et al. Comparison of World Health Organization and Asia-Pacific body mass index classifications in COPD patients. Int J Chron Obstruct Pulmon Dis. 2017;12:2465–7. doi: 10.2147/COPD.S141295. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. The DHS program - wealth-index-construction. Available from: https://dhsprogram.com/topics/wealth-index/Wealth-Index-Construction.cfm. accessed on March 25, 2026.
- 20. Food and Nutrition Technical Assistance (FANTA III). USAID from the American people. Household dietary diversity score (HDDS) for measurement of household food access: Indicator guide (v.2). Available from: https://www.fantaproject.org/sites/default/files/resources/HDDS_v2_Sep06_0.pdf, accessed on April 9, 2026.
- 21.Vickers AJ, Elkin EB. Decision curve analysis: A novel method for evaluating prediction models. Med Decis Making. 2006;26:565–74. doi: 10.1177/0272989X06295361. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Luyckx VA, Tonelli M, Stanifer JW. The global burden of kidney disease and the sustainable development goals. Bull World Health Organ. 2018;96:414–422D. doi: 10.2471/BLT.17.206441. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.GBD Chronic Kidney Disease Collaboration Global, regional, and national burden of chronic kidney disease, 1990-2017: A systematic analysis for the global burden of disease study 2017. Lancet. 2020;395:709–33. doi: 10.1016/S0140-6736(20)30045-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Jha V, Garcia-Garcia G, Iseki K, Li Z, Naicker S, Plattner B, et al. Chronic kidney disease: Global dimension and perspectives. Lancet. 2013;382:260–72. doi: 10.1016/S0140-6736(13)60687-X. [DOI] [PubMed] [Google Scholar]
- 25.Singh AK, Farag YMK, Mittal BV, Subramanian KK, Reddy SRK, Acharya VN, et al. Epidemiology and risk factors of chronic kidney disease in India – results from the SEEK (Screening and early evaluation of kidney disease) study. BMC Nephrol. 2013;14 doi: 10.1186/1471-2369-14-114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Rai PK, Rai P, Bedi S. Prevalence and risk factors of chronic kidney disease in overweight and obese population in a tertiary care hospital in North India. Saudi J Kidney Dis Transpl. 2019;30:1431–8. doi: 10.4103/1319-2442.275488. [DOI] [PubMed] [Google Scholar]
- 27.Biswas P, Sahu AK, Nath SR, Mir SA, Naik PK, Kariali E, et al. Prevalence of chronic kidney disease and its association with pesticide exposure in Bargarh district, Odisha, India. Indian J Nephrol. 2024;34:467–74. doi: 10.25259/ijn_188_23. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Hamilton SA, Jarhyan P, Fecht D, Venkateshmurthy NS, Pearce N, Narayan KMV, et al. Environmental risk factors for reduced kidney function due to undetermined cause in India: an environmental epidemiologic analysis. Environ Epidemiol. 2021;5:e170. doi: 10.1097/EE9.0000000000000170. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Akurathi P, Aligina KK, Rao UP, Thanuja P, Pothukuchi M, Sridevi N. Assessment of prevalence and risk factors of chronic kidney disease in A Konduru tribal region of Andhra Pradesh, India: A cross-sectional study. J Family Med Prim Care. 2025;14:2958–65. doi: 10.4103/jfmpc.jfmpc_347_25. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Devarajan P. The use of targeted biomarkers for chronic kidney disease. Adv Chronic Kidney Dis. 2010;17:469–79. doi: 10.1053/j.ackd.2010.09.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Provenzano M, Rotundo S, Chiodini P, Gagliardi I, Michael A, Angotti E, et al. Contribution of predictive and prognostic biomarkers to clinical research on chronic kidney disease. Int J Mol Sci. 2020;21:5846. doi: 10.3390/ijms21165846. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Boutilier JJ, Chan TCY, Ranjan M, Deo S. Risk stratification for early detection of diabetes and hypertension in resource-limited settings: Machine learning analysis. J Med Internet Res. 2021;23:e20123. doi: 10.2196/20123. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Rashed-Al-Mahfuz M, Haque A, Azad A, Alyami SA, Quinn JMW, Moni MA. Clinically applicable machine learning approaches to identify attributes of chronic kidney disease (CKD) for use in low-cost diagnostic screening. IEEE J Transl Eng Health Med. 2021;9:4900511. doi: 10.1109/JTEHM.2021.3073629. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Supplementary Figure