Abstract
Objectives
To develop predictive models for early and overall tuberculosis (TB) deaths for prospective use at TB diagnosis in resource-constrained TB programme settings.
Design
Statewide cohort study using routinely captured secondary data.
Setting
With the majority of TB deaths being early (within 2 months), India’s TB programme’s information management system (Ni-kshay)-dependent death prediction models (using age, gender, TB site, previous treatment, microbiological confirmation, HIV, diabetes and bank account availability) are not feasible for prospective use, as few variables are captured at diagnosis. Utilising routinely captured triage variables for severe illness at diagnosis (body mass index, pedal oedema, respiratory rate, oxygen saturation and ability to stand without support) from an ongoing statewide and state-specific differentiated TB care initiative to reduce TB deaths in Tamil Nadu state (southern India, 80 million population with 0.1 million annual notifications), robust models for prospective use were developed.
Participants
Adults (aged ≥15 years) with TB (not known to be drug-resistant at diagnosis) that were notified from public facilities of Tamil Nadu from July 2022 to June 2023.
Outcome measures
Early and overall (within 12 months of notification) TB deaths. Area under the receiver operating characteristic curve (AUC) was used to assess accuracy of models built using modified Poisson regression.
Results
Among 55 971 adults, the overall death rate was 7.4%, and 67.9% of the deaths were early. In predicting overall deaths, accuracy of the model using all Ni-kshay variables (AUC 0.716 (95% CI 0.707 to 0.725)) was as good as the model using triage variables for severe illness only (AUC 0.701 (95% CI 0.691 to 0.711)). To the latter, adding potentially capturable Ni-kshay variables at diagnosis (age, gender, TB site, previous treatment and microbiological confirmation) significantly improved model accuracy (AUC 0.754 (95% CI 0.745 to 0.763)). Further addition of remaining Ni-kshay variables did not improve accuracy significantly. Death prediction equations were generated for these models.
Conclusion
Simple and easily measurable triage variables for severe illness should be routinely captured at TB diagnosis. A death prediction calculator (http://44.208.93.99/) based on these variables (specifically triage variables for severe illness combined with age, gender, TB site, previous treatment and microbiological confirmation) may be used by Indian states and high TB burden countries seeking scalable, data-driven interventions to reduce TB deaths.
Keywords: Differentiated TB care, Risk stratification, TB mortality, Predictive models, Triaging for severe illness, Operational Research
Strengths and limitations of this study.
This is a large, statewide cohort study assessing the predictive utility of triage variables for severe illness recorded at the time of tuberculosis (TB) diagnosis in routine programme settings.
The study evaluates multiple hypothetical scenarios (models 2–4) to ensure that the prediction tools are flexible and applicable in resource-constrained real-world settings.
The use of routine programmatic data may introduce potential errors or misclassifications of independent variables.
Approximately 17% of adults with TB who were not triaged at diagnosis were excluded, which may affect the generalisability of the results.
These prediction models were specifically designed for use at the time of diagnosis and cannot be applied to predict mortality at later stages of TB treatment.
Introduction
WHO’s End TB Strategy provides a flexible framework for countries to reduce tuberculosis (TB)-related deaths (henceforth, referred to as TB deaths) by 95% by 2035.1 The progress towards reducing TB deaths could be accelerated by reducing TB incidence, reducing deaths among incident TB (early initiation of appropriate regimen, treatment adherence, severity assessment and comorbidity management), addressing social determinants of health and ensuring universal health coverage.2
In high TB burden countries, considering majority deaths are early (within 2 months of diagnosis, within 8 weeks of intensive phase),3–5 it is imperative to decide on the need for inpatient care (severity assessment) at TB diagnosis before linking the person with TB to a peripheral health facility for ambulatory directly observed treatment (differentiated TB care to reduce TB deaths). Clinical, laboratory and radiological capacities vary across facilities that diagnose TB. Thus, there is a need for death prediction models that are flexible in line with the available resources and capacity, using variables available at diagnosis in resource-constrained programme settings.3
Most TB death prediction models are either from clinical resourceful settings or focused on drug-resistant TB or HIV-associated TB, and they do not cater to drug-susceptible TB in non-HIV settings.6–10 Additionally, models using programme data utilise retrospective data on baseline characteristics, many of which are not capturable at TB diagnosis to take appropriate action.6 Hence, they are not prospectively implementable.
Tamil Nadu is a southern India state with ≈80 million population, ≈0.1 million annual TB notifications and 6.6% TB death rate among all people with TB notified from public facilities in 2021.11 Between April 2022 and December 2025, the state implemented a differentiated TB care initiative to reduce TB deaths called Tamil Nadu Kasanoi Erappila Thittam (TN-KET, meaning TB death-free project in Tamil).12–14 At TB diagnosis, a paper-based TN-KET triage tool captured five variables (body mass index (BMI), pedal oedema, respiratory rate, oxygen saturation and ability to stand without support) to measure five indicators for identification of three conditions (very severe undernutrition, respiratory insufficiency and poor performance status) (Box 1).3 15 These variables are strong predictors of TB deaths,16–19 easy to collect at diagnosis in routine programme settings from primary to tertiary healthcare facilities (TB diagnosis centres)12–14 and were not captured in any other national TB programmes. These were captured in Severe TB Web Application (TB SeWA), the state-specific e-tool of TNKET.12–14 20 As a public health guidance, those with any of the three conditions (triage-positive) were prioritised for inpatient care. This is probably the first such model implemented statewide in routine healthcare settings.12–14
Box 1. Triage tool for severe illness at diagnosis among adults (≥15 years) with TB (without known drug-resistant disease at diagnosis) notified from public facilities in Tamil Nadu, India 3 15*.
If at least one of the following is present, then the person with TB is ‘triage-positive’ (to be prioritised for comprehensive assessment, confirmation of severe illness and inpatient care)
BMI less than or equal to (≤) 14.0 kg/m2#
BMI less than or equal to (≤) 16.0 kg/m2 with pedal oedema#
Respiratory rate more than (>) 24 per minute##
Oxygen saturation less than (<) 94%##
Not able to stand without support (poor performance status - standing with support/squatting/sitting/bedridden)
Ni-kshay (India TB programme’s information management system)-dependent death prediction models may not be feasible for prospective use, as few variables are captured at diagnosis. Availability of triage variables for severe illness at diagnosis in Tamil Nadu provided the opportunity to develop robust and useful death prediction models.14
This study aimed to develop predictive models (applicable at TB diagnosis) for early and overall TB deaths using triage variables from TB SeWA and routinely capturable baseline variables from Ni-kshay. Specific objectives were to (i) assess the improvement in the accuracy of Ni-kshay variables-based death prediction model after incorporating the triage variables and (ii) compare four prediction models (along with their accuracy estimates) in four hypothetical scenarios in line with the availability of resources and capacity at TB diagnosis centres in routine programme settings.
Methods
Study design
This study utilised a cohort design, drawing on secondary data from Ni-kshay and TB SeWA.
Setting
Tamil Nadu national TB elimination programme infrastructure includes 35 districts (30 excluding Chennai), ≈460 subdistrict-level TB units and ≈5250 peripheral health institutions (PHIs), of which ≈2780 are in public sector. Paper-based registers maintained at TB units record notifications of people with TB, management details and treatment outcomes. These details are updated in Ni-kshay by the senior TB treatment supervisor of each TB unit. It is standard practice for the diagnosing TB unit to notify at diagnosis and refer the person with TB to the PHI closest to the residence for ambulatory directly observed treatment. In the current TB unit (where the person is taking treatment), the senior TB treatment supervisor transcribes most of the baseline variables into Ni-kshay, resulting in an induced/ unavoidable delay in capturing baseline variables.
Participants
TN-KET focused on adult (≥15 years) public-notified TB (not known to be drug-resistant at diagnosis) in 30 TB districts of Tamil Nadu (all districts except Chennai, the capital city).14 21 In this study, those notified between July 2022 and June 2023 were included. As per programme guidelines, children (<15 years) with presumptive TB are referred to paediatrician for further workup. People with TB who are found to be drug-resistant are routinely referred for further pre-treatment evaluation to pulmonologists in tertiary care centres. Therefore, these two categories of people with TB were not included under TN-KET.14 21
Scenarios for the TB death prediction models, variables and sources
Models for four hypothetical scenarios were identified (table 1). Model 1 constructed using baseline variables from Ni-kshay reflected the current standard practice, in which TN-KET triage variables are not captured.
Table 1.
Scenarios for TB death prediction models, variables and sources
| Model | Scenario | Ni-kshay variables* | TB SeWA variables |
| 1 | Current standard practice, in which TN-KET triage variables are not captured | Age, gender, test used for diagnosis, TB site, previous treatment, microbiological confirmation, HIV status, diabetes status, availability of bank details, time interval between diagnosis and treatment initiation, level of the diagnosing facility† and diagnosing district† | − |
| 2 | TN-KET triage variables at diagnosis are used for risk assessment in resource-constrained settings (limited diagnostic and clinical capacity) | − | Level of diagnosing facility†, diagnosing district† and triage-related variables (BMI, pedal oedema, respiratory rate, oxygen saturation and ability to stand without support) |
| 3 | Resource-constrained setting where model 2 is supported by potentially capturable variables at diagnosis from Ni-kshay | Model 2 plus age, gender, test used for diagnosis, TB site, previous treatment and microbiological confirmation | Same as model 2 |
| 4 | Near-ideal scenario where all variables from Ni-kshay and TB SeWA are capturable at diagnosis | Model 3 plus HIV status, diabetes status, availability of bank details and time interval between diagnosis and treatment initiation | Same as model 2 |
*Few variables are captured in Ni-kshay at TB diagnosis.
†The level of the diagnosing facility and diagnosing district are capturable in both Ni-kshay and TB SeWA.
BMI, body mass index; TB, tuberculosis; TB SeWA, severe TB web application; TN-KET, Tamil Nadu Kasanoi Erapilla Thittam.
Model 2 is applicable where TN-KET triage variables at diagnosis are used in resource-constrained settings (limited diagnostic and clinical capacity). Model 3 represents a resource-constrained setting where model 2 is supported by potentially capturable variables at diagnosis from Ni-kshay, such as age, gender, diagnostic test used, TB site, previous treatment and microbiological confirmation. Model 4 represents a near-ideal scenario in which all variables from Ni-kshay and TB SeWA are capturable at diagnosis. Model 4 is model 3 plus diabetes status, HIV status, availability of bank account (non-availability proxy for poor socio-economic status) and time interval between diagnosis and treatment initiation from Ni-kshay.
Models 1 and 4 may not be feasible for prospective use as few variables are captured in Ni-kshay at TB diagnosis. Models 3 and 4 can be potentially implemented in Ni-kshay if it also captures triage variables or in TB SeWA if both e-platforms ‘talk’ to each other (through application programming interface integration). Model 3 can be considered if there is an administrative will to capture the potentially capturable variables from Ni-kshay and the five triage variables for severe illness at diagnosis itself.
Statistical analysis
Data analysis was performed using Stata (V.12.1, StataCorp LP, College Station, TX, USA) for the descriptive and analytical components, while R (V.2024.12.1+563) was used for constructing the receiver operating characteristic plots. The authors had access to all the variables from Ni-kshay and TB SeWA that were required for analysis.
Programme-reported early (within 2 months) and overall (routinely reported maximum within 12 months in Ni-kshay) TB deaths were the outcomes. Death was defined as TB-related death (TB may or may not be the underlying cause of death) occurring anytime from notification to outcome declaration (maximum 12 months from notification).22
Given the binary nature of the outcome variable, modified Poisson regression (enter method) with robust variance estimates was used.23 Variables with crude p<0.20 were included in the regression models. Continuous variables (age, BMI, respiratory rate and oxygen saturation) were retained as such in the models without categorisation. For missing continuous variables or implausible outliers in continuous variables, imputation was performed using mean values from the dataset. For handling missing data in categorical variables, separate ‘missing’ categories were retained if the extent of missingness was relatively high (Ni-kshay variables) and imputed as mode if the extent of missingness was relatively low (triage variables).
Modified Poisson regression models were built for each of the four scenarios (models 1–4). Within each scenario, four outcomes were considered: early death, overall death and their respective sensitivity analyses. Sensitivity analysis assumed all programme-reported pre-treatment and during-treatment loss to follow-up (not taking treatment consecutively for 2 months) as TB deaths.22 The predictive performance of each model was evaluated using the area under the receiver operating characteristic curve (AUC). To ensure statistical rigour beyond overlapping 95% CIs, DeLong’s test was employed for formal pairwise comparisons of the AUCs. The primary analysis employed modified Poisson regression with robust variance estimates to determine the factors associated with mortality, reporting results as risk ratios (RR) with 95% CIs. Although modified Poisson regression was used to estimate RR, individual predicted probabilities were derived using the logit transformation. This approach was adopted under the assumption that for outcomes with relatively low frequency, the RR provides a robust approximation for risk estimation, whereas the logit link ensures clinically interpretable probabilities constrained between 0 and 1. Beta coefficients were utilised for creating prediction equations, with the linear predictor (η) as the output. The probability of TB deaths was then calculated using the formula exp(η) / [1+exp(η)].
To ensure the prediction equations remain applicable for prospective use in routine programmatic settings, we adopted a pragmatic approach to handling missing input values. If one or more predictor variables are missing at the time of prospective use at TB diagnosis, continuous variables will get imputed using their mean values and categorical variables using their mode. The pattern of missingness was assessed using Little’s missing completely at random test for the variables respiratory rate, oxygen saturation and BMI. The test yielded χ²(7) = 95.46, p<0.001, indicating that the data were not missing completely at random. The proportion of missing data was low across these variables, <10% for each (for those who were triaged). After correcting data entry errors (replacing 0 and 99), the remaining missing data were handled using mean imputation. This strategy avoids excluding variables from the prediction process and helps maintain the accuracy and reliability of the prediction when deployed by health staff in real-world conditions.24 For model 1, missing programmatic data were retained as a separate category, and no imputation was performed. For models 2–4, a sensitivity analysis using complete case analysis (CCA) was conducted to assess the robustness of the mean/mode imputation strategy. Model performance (AUC) from the CCA was compared against the primary models to ensure consistency (online supplemental table S19).
bmjopen-16-6-s001.pdf (531.9KB, pdf)
Results
Of a total of 55 971 study participants notified, 82.9% (n=46 417) were triaged at diagnosis. The overall death rate was 7.4% (n=4141), with an early death rate of 5.0% (n=2810, 67.9% of overall deaths). The overall death rate based on triage results was as follows: 19.8% (1124/5681) among triage-positive (severely ill), 5.1% (2074/40,736) among triage-negative (not severely ill based on triaging) and 8.8% (843/9554) among not triaged.
Among those triaged, the detailed frequencies of early and overall deaths by baseline characteristics from Ni-kshay and TB SeWA are provided in table 2. Predictive models were built using data from these 46 417 triaged study participants. The test used for diagnosis was removed from the models because of multicollinearity (variance inflation factor >10). Time interval between diagnosis and treatment initiation was excluded from the models (crude p>0.20). Model diagnostics and beta coefficients for the variables included in models 1–4 are provided in online supplemental tables S3–S18.
Table 2.
Baseline characteristics and mortality outcomes among public notified adults with TB (not known to be drug-resistant at diagnosis) in Tamil Nadu, India, July 2022–June 2023 (n=46 417)
| Baseline characteristics | Total | Early deaths | Overall deaths | |||
| n | (%)* | n | (%)† | n | (%)† | |
| Total | 46 417 | (100.0) | 2179 | (4.7) | 3198 | (6.9) |
| Variables from Ni-kshay | ||||||
| Age, years | ||||||
| 15–44 | 17 858 | (38.7) | 454 | (2.5) | 697 | (3.9) |
| 45–64 | 21 644 | (46.4) | 1129 | (5.2) | 1657 | (7.7) |
| ≥ 65 | 6915 | (14.9) | 596 | (8.6) | 844 | (12.2) |
| Gender | ||||||
| Women | 12 666 | (27.4) | 437 | (3.5) | 650 | (5.1) |
| Men | 33 741 | (72.6) | 1742 | (5.1) | 2548 | (7.6) |
| Transgender | 10 | (0.02) | 0 | (0) | 0 | (0) |
| Bank details available | ||||||
| Yes | 45 319 | (96.2) | 1889 | (4.2) | 2885 | (6.4) |
| No | 1098 | (3.8) | 290 | (26.4) | 313 | (28.5) |
| Test used for diagnosis | ||||||
| Rapid molecular tests | 7923 | (18.9) | 385 | (4.9) | 576 | (7.3) |
| Microscopy/culture | 25 334 | (52.5) | 1149 | (4.5) | 1701 | (6.7) |
| Chest radiograph | 4904 | (10.9) | 242 | (4.9) | 350 | (7.1) |
| Others/missing | 8256 | (17.7) | 403 | (4.9) | 571 | (7.0) |
| Microbiological confirmation | ||||||
| Yes | 34 916 | (75.2) | 1609 | (4.6) | 2380 | (6.8) |
| No | 11 501 | (24.8) | 570 | (4.9) | 818 | (7.1) |
| Site | ||||||
| Pulmonary | 38 747 | (82.0) | 1739 | (4.5) | 2630 | (6.8) |
| Extrapulmonary | 7265 | (16.0) | 302 | (4.1) | 428 | (5.9) |
| Missing | 405 | (2.1) | 138 | (34.1) | 140 | (34.6) |
| Previous treatment | ||||||
| Yes | 5359 | (11.5) | 233 | (4.4) | 392 | (7.3) |
| No | 41 058 | (88.5) | 1946 | (4.7) | 2806 | (6.8) |
| HIV | ||||||
| Negative | 44 491 | (95.1) | 1941 | (4.4) | 2871 | (6.5) |
| Positive | 1768 | (4.1) | 189 | (10.7) | 276 | (15.6) |
| Unknown | 158 | (0.8) | 49 | (31.0) | 51 | (32.3) |
| DM | ||||||
| No | 33 916 | (73.0) | 1560 | (4.6) | 2299 | (6.8) |
| Yes | 12 198 | (25.6) | 550 | (4.5) | 826 | (6.8) |
| Unknown | 303 | (1.4) | 69 | (22.8) | 73 | (24.1) |
| Time interval (in days) between diagnosis and treatment initiation | ||||||
| ≤1 | 28 088 | (60.5) | 1302 | (4.6) | 1907 | (6.7) |
| 2–6 | 14 604 | (31.5) | 609 | (4.1) | 939 | (6.4) |
| ≥7 | 3322 | (7.1) | 130 | (3.9) | 212 | (6.3) |
| Missing | 403 | (0.9) | 138 | (34.2) | 140 | (34.7) |
| Triage-related indicators/variables | ||||||
| Level of diagnosing facility | ||||||
| Tertiary | 20 420 | (44.4) | 1193 | (5.8) | 1701 | (8.3) |
| Secondary | 17 973 | (38.3) | 739 | (4.1) | 1131 | (6.3) |
| Primary | 7996 | (17.2) | 247 | (3.1) | 366 | (4.6) |
| Missing | 28 | (0.1) | – | – | ||
| BMI ≤14 kg/m2 | ||||||
| Yes | 2331 | (5.1) | 351 | (15.1) | 475 | (20.4) |
| No | 44 086 | (94.9) | 1828 | (4.1) | 2723 | (6.1) |
| BMI 14.1–16 kg/m2 with bilateral pedal oedema | ||||||
| Yes | 156 | (0.3) | 36 | (23.1) | 42 | (26.9) |
| No | 46 261 | (99.7) | 2143 | (4.6) | 3156 | (6.8) |
| Respiratory rate >24 | ||||||
| Yes | 2238 | (4.8) | 332 | (14.9) | 419 | (18.7) |
| No | 44 179 | (95.2) | 1847 | (4.2) | 2779 | (6.3) |
| SpO2<94% | ||||||
| Yes | 1116 | (2.4) | 207 | (18.6) | 259 | (23.2) |
| No | 45 301 | (97.6) | 1972 | (4.4) | 2939 | (6.5) |
| Unable to stand without support | ||||||
| Yes | 1865 | (4.0) | 463 | (24.8) | 566 | (30.3) |
| No | 44 552 | (96.0) | 1716 | (3.9) | 2632 | (5.9) |
| Classification based on conditions | ||||||
| Very severe undernutrition‡ | ||||||
| Yes | 2487 | (5.4) | 387 | (15.6) | 517 | (20.8) |
| No | 43 930 | (94.6) | 1792 | (4.1) | 2681 | (6.1) |
| Respiratory insufficiency‡ | ||||||
| Yes | 2855 | (6.2) | 437 | (15.3) | 552 | (19.3) |
| No | 43 562 | (93.9) | 1742 | (4.0) | 2646 | (6.1) |
| Poor performance status‡ | ||||||
| Yes | 1860 | (4.0) | 463 | (24.9) | 566 | (30.4) |
| No | 44 557 | (96.0) | 1716 | (3.9) | 2632 | (5.9) |
| Very severe undernutrition‡ + Respiratory insufficiency‡ | ||||||
| Yes | 610 | (1.3) | 160 | (26.2) | 195 | (32.0) |
| No | 45 807 | (98.7) | 2019 | (4.4) | 3003 | (6.6) |
| Very severe undernutrition‡ + poor performance status‡ | ||||||
| Yes | 362 | (0.8) | 129 | (35.6) | 158 | (43.7) |
| No | 46 055 | (99.2) | 2050 | (4.5) | 3040 | (6.6) |
| Respiratory insufficiency‡ + poor performance status‡ | ||||||
| Yes | 638 | (1.4) | 185 | (29.0) | 224 | (35.1) |
| No | 45 779 | (98.6) | 1994 | (4.4) | 2974 | (6.5) |
| Very severe undernutrition‡ + respiratory insufficiency‡ + poor performance status‡ | ||||||
| Yes | 155 | (0.3) | 65 | (42.0) | 74 | (47.7) |
| No | 46 262 | (99.7) | 2114 | (4.6) | 3124 | (6.8) |
*Column percentages.
†Row percentages.
‡Very severe undernutrition=either positive for BMI ≤ 14 kg/m2 or 14.1–16 kg/m2 with bilateral pedal oedema. Respiratory insufficiency=either positive for RR > 24/min or Spo2 < 94%. Poor performance status=inability to stand without support.
BMI, body mass index; DM, diabetes mellitus; TB, tuberculosis; TB SeWA, severe TB web application; TN-KET, Tamil Nadu Kasanoi Erapilla Thittam.
Improvement in model 1 after incorporating variables used for triaging
For early deaths, model 1 demonstrated AUC of 0.732 (95% CI 0.721 to 0.742), which significantly increased to 0.787 (95% CI 0.778 to 0.797) when triage variables were added. For overall deaths, model 1 demonstrated AUC of 0.716 (95% CI 0.707 to 0.725), which also significantly increased to 0.767 (95% CI 0.759 to 0.776) when triage variables were added (online supplemental figure S1). The results were consistent after the sensitivity analyses (table 3).
Table 3.
Area under the curve (AUC) of the TB death rate prediction models with early and overall deaths as outcomes among public notified adults with TB (not known to be drug-resistant at diagnosis) in Tamil Nadu, India, July 2022–June 2023 (n=46 417)
| Early deaths | Overall deaths | |||||||
| Without LTFU | With LTFU | Without LTFU | With LTFU | |||||
| AUC | 95% CI | AUC | 95% CI | AUC | 95% CI | AUC | 95% CI | |
| Model 1 | 0.732 | 0.721 to 0.742 | 0.734 | 0.724 to 0.744 | 0.716 | 0.707 to 0.725 | 0.694 | 0.686 to 0.703 |
| Model 2 | 0.715 | 0.703 to 0.726 | 0.688 | 0.677 to 0.699 | 0.701 | 0.691 to 0.711 | 0.671 | 0.662 to 0.680 |
| Model 3 | 0.772 | 0.762 to 0.783 | 0.752 | 0.742 to 0.762 | 0.754 | 0.745 to 0.763 | 0.712 | 0.704 to 0.720 |
| Model 4 | 0.787 | 0.778 to 0.797 | 0.778 | 0.769 to 0.788 | 0.767 | 0.759 to 0.776 | 0.732 | 0.724 to 0.740 |
LTFU, loss to follow-up; TB, tuberculosis.
Comparing all four models
Model 2 achieved an AUC of 0.715 (95% CI 0.703 to 0.726) for early deaths and 0.701 (95% CI 0.691 to 0.711) for overall deaths. Its predictive performance was not significantly different from model 1. When compared model 2, model 3 had significantly better predictive performance: AUC of 0.772 (95% CI 0.762 to 0.783) for early deaths and 0.754 (95% CI 0.745 to 0.763) for overall deaths. When compared with model 3, model 4 did not have significantly better predictive performance: AUC of 0.787 (95% CI 0.778 to 0.797) for early deaths and 0.767 (95% CI 0.759 to 0.776) for overall deaths (online supplemental figure S1). Pairwise comparisons using DeLong’s test demonstrated that the discriminative accuracy (AUC) of models 1–4 differed significantly from one another (p<0.001).
Furthermore, sensitivity analysis using CCA confirmed the robustness of our missing data handling strategy; the predictive accuracy across all models remained consistent with the primary findings (online supplemental table S19), suggesting that the pragmatic mean/mode imputation did not introduce significant bias. Beyond AUC, comprehensive performance metrics including sensitivity, specificity and accuracy for all models across both early and overall mortality outcomes are detailed in online supplemental table S20).
Tuberculosis death prediction formulae
The predictive equations generated from all four models are provided in online supplemental table S1. For ease of demonstrating the predicted death rates, model 2 was modified in which the three triage conditions (categorical variables) were included instead of triage variables (model data not shown). The predicted death rates for each of the three conditions and their combinations are depicted in table 4 after averaging out across districts and levels of diagnosing facility. Individually, in the absence of the other two conditions, the following were the respective predicted early and overall TB death rates: poor performance status (16.2% and 18.1%), very severe undernutrition (10.1% and 13.0%) and respiratory insufficiency (8.4% and 10.3%). The predicted death rates among those with any two conditions were highest for the combination of poor performance status and very severe undernutrition (31.6% for early and 33.9% for overall death rate).
Table 4.
Predicted TB death rate (model 2 was modified in which three triage conditions (categorical variables) were included instead of triage variables) based on the triage condition at diagnosis among adults with TB (not known to be drug-resistant at diagnosis)
| Triage condition at TB diagnosis | Predicted early TB death rate (%)* | Predicted overall TB death rate (%)* |
| All three conditions absent (triage-negative) | 2.2 | 3.9 |
| Very severe undernutrition† only | 10.1 | 13.0 |
| Respiratory insufficiency† only | 8.4 | 10.3 |
| Poor performance status† only | 16.2 | 18.1 |
| Very severe undernutrition and respiratory insufficiency | 18.2 | 21.1 |
| Very severe undernutrition and poor performance status | 31.6 | 33.9 |
| Respiratory insufficiency and poor performance status | 27.4 | 28.4 |
| All three conditions present | 47.2 | 47.7 |
*The predicted death rate varies across the level of the diagnosing facility (primary/secondary/tertiary) and the 30 districts, and this was averaged out.
†Very severe undernutrition=either positive for body mass index ≤ 14 (or) 14.1–16 with bilateral pedal oedema. Respiratory insufficiency=either positive for respiratory rate > 24 or oxygen saturation < 94%. Poor performance status=inability to stand without support.
TB, tuberculosis.
Discussion
Key findings
This is the first study to assess the predictive utility of BMI, pedal oedema, respiratory rate, oxygen saturation and ability to stand without support (recorded at diagnosis in Tamil Nadu, southern India) for TB deaths in routine programmatic settings. There were three key findings. First, the addition of these triage variables to a TB death prediction model based on all baseline variables captured by India’s TB programme (Ni-kshay) significantly improved the prediction accuracy. Second, by themselves, the accuracy of TB death prediction based on triage variables was similar to that based on all baseline variables captured by India’s TB programme. Third, the addition of potentially capturable variables at diagnosis from Ni-kshay significantly improved the death prediction based on triage variables only. However, further addition of variables from Ni-kshay did not improve accuracy.
Strengths and limitations
The analysed cohort is highly representative of the adult drug-susceptible TB population seeking care in the public sector across Tamil Nadu. This is supported by the statewide coverage of 30 districts, a large sample size of 55 971 participants and a high triaging rate of 82.9% within routine programme settings.
Considering 70% of deaths were early, assessing three scenarios (models 2–4) at diagnosis in line with the availability of resources and capacity makes it suitable for real-world application. A recent systematic review highlighted that most TB outcome prediction models were developed using retrospective clinical data, with limited external validity and a lack of integration into routine public health programmes.6 The AUC values in this study were also comparable with those reported in the review, demonstrating strong discriminatory power and suggesting that programme-linked prediction models can be both scalable and effective in real-world settings.6
There were four limitations. First, 17% of adults with TB who were not triaged were excluded from the prediction models. When compared with the triaged group, previous analysis has shown that those not triaged were more likely to have extrapulmonary TB, diagnosed in high notification districts, diagnosed in teaching hospitals and transferred out of the district.14 Second, these prediction models can be used only at the time of diagnosis and cannot be used later for predicting TB deaths. Third, the use of routine programme data introduces potential errors or misclassifications. We see no reasons for differential errors in data capture of triage and Ni-kshay variables among those who experience and do not experience death. The results were consistent after the sensitivity analysis assuming all programme-reported pre-treatment and during-treatment losses (not taking treatment consecutively for 2 months) as TB deaths. Hence, we do not expect the misclassification of any of the independent variables to be differential.
Finally, lack of external validation is another limitation. As this is the first statewide programme to capture these specific triage variables at diagnosis, no independent dataset was available for comparison. Future studies should evaluate the performance of these models in different regional contexts to ensure wider generalisability.
Policy and practice relevance
The prediction equations are intended to be integrated into digital TB surveillance systems (such as Ni-kshay or TB SeWA), and it has been designed for use by frontline health staff, such as senior treatment supervisors, without requiring advanced clinical or statistical expertise. In addition, a web-based death prediction calculator has been developed to allow any health staff to easily access all 4 models and estimate the predicted probability of TB death (http://44.208.93.99/). It is free to use after appropriate training.25
Even with triage variables used in TN-KET, the death predictive performance was good and comparable to all baseline variables in Ni-kshay. In Tamil Nadu, triage variables are routinely captured (and ongoing) at TB diagnosis within a day.14 These are easy to measure and infer in routine programme settings.14 15 26 Conversely, all baseline variables of Ni-kshay are not feasible to capture at diagnosis. The five triage variables, namely BMI, pedal oedema, respiratory rate, oxygen saturation and ability to stand without support, should be included as baseline variables at TB diagnosis by national TB programmes and acted on before sending the person to a PHI for ambulatory treatment (as done in TN-KET).12–14
Even though the predictive performance of triage variables was good, the addition of potentially capturable baseline variables from Ni-kshay at diagnosis significantly improved the accuracy of prediction. However, incorporating other Ni-kshay variables that are potentially not capturable at the time of diagnosis did not substantially improve the model. The addition of variables not immediately capturable at diagnosis adds complexity without meaningful benefit; these may be deprioritised for triaging and prediction purposes. A simple, triage variables-based model supported by some variables from Ni-kshay may thus be sufficient for prospective real-time implementation. If there is administrative will to capture potentially capturable variables from Ni-kshay at diagnosis (and not in PHI where treatment is provided), then model 3 can be considered in Ni-kshay if it also captures triage variables or in TB SeWA if both e-platforms ‘talk’ to each other through application programming interface integration. Incorporating a machine learning model to transform these static models into a dynamic one will facilitate continuous refinement of the prediction on a regular basis as new cohort data becomes available. Until this happens, the web-based death prediction calculator (http://44.208.93.99/) may be used.25
Conclusion
BMI, pedal oedema, respiratory rate, oxygen saturation and ability to stand without support are not routinely captured by TB programmes. The availability of these five triage variables related to severe illness at TB diagnosis in Tamil Nadu allowed us to generate robust and feasible TB death prediction models. Other states in India and high TB burden countries should also consider including these variables in their existing TB information management systems. These triage variables with (model 3) or without (model 2) other potentially capturable variables at diagnosis (age, gender, TB site, previous treatment and microbiological confirmation) should be used to predict deaths. These models are flexible in line with the available resources and capacity and can therefore be applied in programme settings. The integration of model 2 or 3 into existing digital platforms like TB SeWA or Ni-kshay hold the potential for real-time use at diagnosis, enabling faster decision-making (referral of severely ill for inpatient care) to reduce TB deaths. Until the integration happens, model 3 in web-based death prediction calculator may be used by health staff after appropriate training.
Supplementary Material
Acknowledgments
The authors thank S Gurunathan, ex-Director of Medical and Rural Health Services, and R Narayanababu, ex-Director of Medical Education, Tamil Nadu, for their support in starting TN-KET; the district TB programme staff (district TB officers, medical officers, district programme coordinators, treatment officers, senior treatment supervisors, senior TB laboratory supervisors, health educators, laboratory technicians and TB health visitors among others) for their commitment and support in successfully implementing TN-KET; the nodal physicians from Directorate of Medical Education and Directorate of Medical and Rural Health Services, Tamil Nadu; the following TN-KET investigators: S Lokesh, K Jeyashree, M Kalyanasundaram, S Devika, J Chadwick, G Kiruthika, T Daniel Rajasekar, S Kiran Pradeep, P Ravichandran, D Chokkalingam, Varsha Rajan (from ICMR-NIE); S Aarthi, N Tharageshwari, K Rambha, S M Anandrajkumar, (Tamil Nadu State TB Cell); P K Bhavani (ICMR-NIRT), C Padmapriyadarsini (ex Director, ICMR-NIRT); K.V. Suma, R Parthasarathy, D.P. Pathinathan, D. Chidambaram, M Bhavani Nivetha, J G Thampi, R Janani, L Janani, K Sivagami, R. Ramachandran (Office of the WHO Representative to India); T S Laux (Tufts Medical Center, Boston, USA), S Viswanathan (Jawaharlal Institute of Postgraduate Medical Education & Research, Puducherry); R Sridhar (Chettinad Academy of Research & Education, Kelambakkam); K Krishnamoorthy (Tirunelveli Medical College Hospital, Tirunelveli); M Sakthivel, S Karunakaran, S Rajkumar, M Ramachandran, K D Kanagaraj, M Kaleeswari, V P Durai, R Saravanan, A Sugantha, S Z H M Khan, P Sangeetha, R Vasudevan, R Nedunchezhian, M Sankari, N Jeevanandam, S Ganapathy, V Rajasekaran, T Mathavi, A R Rajaprakash, L Murali, U Pugal, K Sundaralingam, S Savithri, S Vellasamy, D Dheenadayal, P Ashok, K Jayasree, R Sudhakar, K P Rajan (Directorate of Medical and Rural Health Services, Government of Tamil Nadu) and T. S. Selvavinayagam (Directorate of Public Health and Preventive Medicine) for their contribution. Tamil Nadu State TB Cell and the ICMR-NIE jointly lead TN-KET with support from ICMR National Institute for Research in Tuberculosis (ICMR-NIRT) and the WHO Country Office in India. This is facilitated by the National Health Mission Tamil Nadu; Directorate of Medical and Rural Health Services, Tamil Nadu; Directorate of Medical Education, Tamil Nadu; and the Directorate of Public Health and Preventive Medicine, Tamil Nadu. TN-KET (and the operational research around it) is being implemented as a Tamil Nadu health system initiative in routine setting using existing resources and workforce, without any study-specific funding.
Footnotes
Contributors: SS, HDS and RS contributed equally as primary authors. SS, HDS and RS conceived and designed the study. AF, RSa, DB, JMM, KG, RV, AJ, DK, EM, TB and MVM were involved in data collection and project implementation. SS, RS and PH performed the statistical analysis. SS, HDS, RS and PH interpreted the data and drafted the manuscript. All authors reviewed and approved the final version of the manuscript. Guarantor: HDS accepts full responsibility for the work and the conduct of the study, had access to the data and controlled the decision to publish. AI has been used for enhancing the readability of the manuscript and not for data analysis or interpretation or any other purpose. Findings and conclusions in this manuscript are those of the authors and do not necessarily represent the official position of their affiliated organisations.
Funding: TN-KET (and the operational research around it) was implemented as a Tamil Nadu health system initiative in a routine setting using existing resources and workforce. The authors received no specific external grant for this research from any funding agency in the public, commercial or not-for-profit sectors. Article processing charges were supported by ICMR-NIE, Chennai.
Competing interests: None.
Patient and public involvement: Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Ethics statement :We received ethics approval from the Institutional Human Ethics Committee of the ICMR National Institute of Epidemiology (ICMR-NIE), Chennai, India (NIE/IHEC/202201-09). As the study was conducted in routine health system settings involving routinely collected secondary data, the need for written informed consent was waived by the ethics committee.
Provenance and peer review: Not commissioned; externally peer reviewed.
Supplemental material: This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.
Data availability statement
Data are available upon reasonable request.
Data availability statement
Data used in this study can be requested from the corresponding author.
Ethics statements
Patient consent for publication
Not applicable.
References
- 1.World Health Organization . The end TB strategy. World Health Organization. Geneva. Report no: WHO/HTM/TB/2015.19. 2015.
- 2.Bhargava A, Bhargava M, Pai M. Tuberculosis: a biosocial problem that requires biosocial solutions. The Lancet 2024;403:2467–9. 10.1016/S0140-6736(24)00489-6 [DOI] [PubMed] [Google Scholar]
- 3.Bhargava A, Bhargava M. Tuberculosis deaths are predictable and preventable: Comprehensive assessment and clinical care is the key. J Clin Tuberc Other Mycobact Dis 2020;19:100155. 10.1016/j.jctube.2020.100155 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Waitt CJ, Peter K Banda N, White SA, et al. Early deaths during tuberculosis treatment are associated with depressed innate responses, bacterial infection, and tuberculosis progression. J Infect Dis 2011;204:358–62. 10.1093/infdis/jir265 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Birlie A, Tesfaw G, Dejene T, et al. Time to Death and Associated Factors among Tuberculosis Patients in Dangila Woreda, Northwest Ethiopia. PLoS One 2015;10:e0144244. 10.1371/journal.pone.0144244 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Peetluk LS, Ridolfi FM, Rebeiro PF, et al. Systematic review of prediction models for pulmonary tuberculosis treatment outcomes in adults. BMJ Open 2021;11:e044687. 10.1136/bmjopen-2020-044687 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Nguyen DT, Jenkins HE, Graviss EA. Prognostic score to predict mortality during TB treatment in TB/HIV co-infected patients. PLoS One 2018;13:e0196022. 10.1371/journal.pone.0196022 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Baluku JB, Apolot PS, Namanda B, et al. A predictive score for early in-patient tuberculosis mortality: A case-control study. J Clin Tuberc Other Mycobact Dis 2024;37:100487. 10.1016/j.jctube.2024.100487 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Ismail N, Kartini K, Abd-Razak A, et al. The mortality predictive model for tuberculosis patients in Malaysia: an application of surveillance data 2013–2014. Int J Infect Dis 2019;79:102. 10.1016/j.ijid.2018.11.254 [DOI] [Google Scholar]
- 10.Krishnamoorthy Y, Ezhumalai K, Murali S, et al. Development of prognostic scoring system for predicting 1-year mortality among pulmonary tuberculosis patients in South India. J Public Health (Oxf) 2023;45:e184–95. 10.1093/pubmed/fdac087 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Ministry of Health and Family Welfare, Central TB Division. India TB report 2023. New Delhi. 2023.
- 12.Shewade HD, Frederick A, Suma KV, et al. Differentiated TB care: Tamil Nadu’s achievements, plans and implications for national TB programmes. Int J Tuberc Lung Dis 2024;28:160–2. 10.5588/ijtld.23.0297 [DOI] [PubMed] [Google Scholar]
- 13.Shewade HD. Standard Operating Procedure for Tamil Nadu Kasanoi Erappila Thittam (TN-KET). Version 6. figshare. Published Online First 2025. 10.6084/M9.FIGSHARE.24564403.V5 [DOI] [Google Scholar]
- 14.Shewade HD, Frederick A, Kiruthika G, et al. The First Differentiated TB Care Model From India: Delays and Predictors of Losses in the Care Cascade. Glob Health Sci Pract 2023;11:e2200505. 10.9745/GHSP-D-22-00505 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Shewade HD, Nagaraja SB, Vanitha B, et al. Screening for Severe Illness at Diagnosis has the Potential to Prevent Early TB Deaths: Programmatic Experience from Karnataka, India. Global Health Sci Pract 2022. 10.9745/GHSP-D-21-00736/-/DCSUPPLEMENTAL [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Lai HH, Lai YJ, Yen YF. Association of Body Mass Index with Timing of Death during Tuberculosis Treatment. PLoS One 2017;12:e0170104. 10.1371/journal.pone.0170104 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.World Health Organization . Management of Severe Malnutrition: A Manual for Physicians and Other Senior Health Workers. Geneva: World Health Organization, 2000. [Google Scholar]
- 18.de Vallière S, Barker RD. Poor performance status is associated with early death in patients with pulmonary tuberculosis. Trans R Soc Trop Med Hyg 2006;100:681–6. 10.1016/j.trstmh.2005.09.007 [DOI] [PubMed] [Google Scholar]
- 19.Zachariah R, Spielmann MP, Harries AD, et al. Moderate to severe malnutrition in patients with tuberculosis is a risk factor associated with early death. Trans R Soc Trop Med Hyg 2002;96:291–4. 10.1016/s0035-9203(02)90103-3 [DOI] [PubMed] [Google Scholar]
- 20.Shewade HD, Frederick A, Kalyanasundaram M, et al. Triaging for Severe Illness amongst Adults with Tuberculosis Followed by Referral and Inpatient Care: A Statewide Pilot in Tamil Nadu, India. Prev Med 2024;1:94–101. 10.4103/PMRR.PMRR_33_23 [DOI] [Google Scholar]
- 21.Shewade HD, Frederick A, Kalyanasundaram M, et al. --Eleven tips for operational researchers working with health programmes: our experience based on implementing differentiated tuberculosis care in south India. Glob Health Action 2023;16:2161231. 10.1080/16549716.2022.2161231 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.World Health Organization . Definitions and reporting framework for tuberculosis – 2013 revision (updated December 2014 and January 2020). Geneva, 2013. [Google Scholar]
- 23.Zou G. A modified poisson regression approach to prospective studies with binary data. Am J Epidemiol 2004;159:702–6. 10.1093/aje/kwh090 [DOI] [PubMed] [Google Scholar]
- 24.Moons KGM, Altman DG, Reitsma JB, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 2015;162:W1–73. 10.7326/M14-0698 [DOI] [PubMed] [Google Scholar]
- 25.ICMR National Institute of Epidemiology (ICMR-NIE) Chennai . Predicting TB deaths supported by triage variables for severe illness at diagnosis. 2025. Available: http://44.208.93.99/ [Accessed 04 June 2026].
- 26.Patel BH, Shewade HD, Davara KB, et al. Screening adults with tuberculosis for severe illness at notification: programme experience from Gujarat, India. Trans R Soc Trop Med Hyg 2022;116:1172–80. 10.1093/trstmh/trac060 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
bmjopen-16-6-s001.pdf (531.9KB, pdf)
Data Availability Statement
Data are available upon reasonable request.
Data used in this study can be requested from the corresponding author.