Development and Validation of a Prognostic Model and Bedside Score for the Neurological Outcome in Children with Tuberculous Meningitis

ABSTRACT.

Only a few studies have explored prognostic factors for tuberculous meningitis (TBM) in children, and an easily applicable bedside prognostic score for TBM has not been developed yet. We compared the sociodemographic, clinical, radiological, and cerebrospinal fluid parameters in the cohort of 94 TBM cases aged 1 to 18 years, with at least 6 months of completed follow-up and determined the prognostic factors associated with poor functional outcome. We assessed our proposed prognostic model using both discrimination and calibration and subsequently used the bootstrap method to validate the model internally. We finally derived an easily applicable bedside prognostic score by rounding off the regression coefficients to the nearest integers. A total of 39 (41%) and 55 (59%) patients had poor and good functional outcomes, respectively, at the end of 6 months (12 died, 13%). In multivariate analysis, a high baseline Pediatric Cerebral Performance Category (PCPC) score, brain infarction in neuroimaging, tonic motor posturing, younger age, and stage III TBM were independent predictors of poor functional outcomes. The final model showed good discrimination (area under the curve = 88.2%, P < 0.001) and good calibration (Hosmer–Lemeshow test, P = 0.53). Bootstrapping also confirmed the internal validity of this model. The PITAS (PCPC score [P], brain infarction in neuroimaging [I], tonic motor posturing [T], age [A], and stage of TBM [S]) score developed from this model has a score ranging from 0 to 12, with a higher score predicting a higher risk of poor functional outcome. The PITAS score performed better than medical research council staging alone in predicting poor functional outcomes (area under the curve = 87.1% versus 82.3%). Our study’s PITAS score, developed and internally validated, has good sensitivity and specificity in predicting poor functional outcomes in pediatric TBM cases at 6 months.

INTRODUCTION

Tuberculous meningitis (TBM) is still the most common cause of chronic meningitis in low- to middle-income countries (LMIC).¹ The yearly burden of morbidity and mortality due to TBM cases in children is high.^2,3 Young children are more likely to develop TBM and disseminated tuberculosis.⁴ Recently, there has been upgradation in the guidelines for administering antituberculosis therapy (ATT).⁵ Most centers in LMICs currently practice surgical interventions such as ventriculoperitoneal shunt for hydrocephalus, when clinically indicated.⁶ Improved compliance with ATT due to stringent government-sponsored programs and the almost universal use of corticosteroids in TBM cases also has contributed to improvement in functional outcome.⁷ Despite all these recent advances, a significant subset of these patients either die or remain with a significant disability even after 6 to 12 months of establishing the diagnosis.⁸ Several investigators have developed various prognostic models in adults with TBM.^9,10 For children, authors from different parts of the world have described clinical features and tried to determine poor prognostic factors.^11,12 A reliable prognostic model in pediatric TBM cases can help clinicians in counseling caregivers at the initiation of therapy regarding future prognosis and identifying the cases likely to have a significant residual disability. In the future, this can guide clinicians in modifying the treatment regimen for this particular subset of patients. Keeping these things in mind, we tried to develop a prognostic model and simple bedside score for the same purpose from our cohort of pediatric TBM cases.

METHODS

We developed and validated the prognostic score using data of children aged 1 to 18 years with TBM treated at our center between August 2019 and January 2022. We began scrutinizing the initial data for duplicate entries by checking with at least three unique identifiers: the name, age, and Universal Health Identity Number provided at our institute. We included only patients who had completed at least 6 months of ATT or died before 6 months, with adequate clinical information, follow-up, and sociodemographic details. To diagnose TBM, we relied on the consensus case definition given by Marais et al.¹³ We excluded those children who had prior major illnesses involving various organ systems (unrelated to the TBM or even predisposed before its development) and subsequently developed TBM. All the cases were diagnosed and managed following Indian extrapulmonary tuberculosis (INDEX-TB) guidelines⁷ and standard recommendations by the Central Tuberculosis Division, Ministry of Health and Family Welfare, Government of India, regarding treating tuberculosis in the pediatric population.

At baseline, each child underwent detailed history-taking, clinical examination, and microbiological and radiological workup, including cerebrospinal fluid (CSF) examination and neuroimaging. Treatment was initiated with ATT and corticosteroids with the addition of antiseizure medications, cerebral antiedema measures, or aspirin as required. Neurosurgical interventions, including ventriculoperitoneal shunt insertion or external ventricular drainage, were performed per clinical indications. At baseline, the severity of TBM was determined according to the UK Medical Research Council (MRC) classification¹⁴ (stages 1–3). Functional status at baseline, at 1-month intervals up to 3 months, then at every 2 to 3 monthly interval until the end of treatment was determined using the Pediatric Cerebral Performance Category (PCPC) and the Functional Status Scale scale.¹⁵ However, our study determined the final outcome at 6 months (earlier for patients who died before). Children in categories 4 (severe disability), 5 (coma or vegetative state), and 6 (death) were considered to have poor functional status, and the rest were considered to have good functional status. We used survival with poor functional status at 6 months as key outcome.

Apart from sociodemographic and clinical examination, the following variables were also noted: results of hematological and biochemical investigations, CSF protein, sugar, total and differential cell count, Ziehl–Neelsen (ZN) staining, culture and polymerase chain reaction/cartridge-based nucleic acid amplification techniques (CB-NAAT), other microbiological investigations (Gram staining, India ink staining, virology, and bacterial and fungal cultures), and contrast-enhanced neuroimaging findings (presence of meningeal enhancement, basal exudates, tuberculomas, hydrocephalus, and infarcts).

Baseline sociodemographic and clinical variables, neuroimaging variables, CSF parameters, and findings of microbiological investigations were treated as potential predictors. Among the clinical and sociodemographic variables, age of onset, age at presentation, gender, number of siblings and family members, nutritional status (presence and severity of underweight, wasting, and stunting), micronutrient deficiency (especially nutritional anemia), immunization status (especially for bacillus Calmette-Guerin), residence (urban/rural), domestic overcrowding, socioeconomic status of family, presence of an active contact with tuberculosis (TB) case in family or community, presence of contact with a multidrug resistant TB case, common symptoms of TBM such as duration and maximum grade of fever, subjective grading of severity of headache by caregiver rated by a numeric rating scale (in conscious patients), severity of poor sensorium (measured by Glasgow Coma Scale [GCS] at presentation), seizures (semiology, frequency, occurrence of status epilepticus, number of antiseizure medications required to achieve seizure control), presence of papilledema, ocular motor nerves involvement, presence and severity of focal neurological deficits, disease severity and disability, and clinical signs of brain herniation (including decerebrate and decorticate posturing, Cushing triad, unequal pupil, jerky breathing) were considered potential predictors.

Neuroimaging variables included meningeal enhancement, hydrocephalus, presence of periventricular ooze, basal exudates, and infarcts. Laboratory parameters included in the analysis were CSF opening pressure, protein, sugar, cell count, HIV status, and drug resistance.

Statistical analysis.

Statistical analysis was performed using the software Statistical Package for Social Sciences—version 29 (SPSS-29, IBM, Chicago, IL). Continuous variables were presented as mean/SD or median/interquartile range [IQR], depending on whether the variable is normally distributed (Z score of skewness within ± 3.29). Categorical variables were presented as frequency (in percentage) with a 95% confidence interval (CI). All the possible aforementioned predictors were initially subjected to univariate analysis to shortlist the baseline predictors associated with poor functional outcomes at 6 months. A χ² test/Fisher’s exact test was chosen for categorical variables, whereas an independent sample t test/Mann–Whitney U test was used for continuous variables in this univariate analysis. Those with P values < 0.05 were considered significant predictors. These were subsequently subjected to multivariate binary logistic regression analysis to determine independent predictors.

We decided to adapt a stepwise backward elimination procedure if the number of independent predictors exceeded the maximum number of candidate predictors allowed for our sample population. For this, we first assumed that, for our study purpose, we would allow a maximum of that number of predictors in the proposed score, where there will be at least 10 participants with dependent or outcome variables per candidate predictor. We determined this event per predictor variable number based on the methods described by Bezda-Delgado et al.¹⁶

We began with a full model and then applied a stepwise backward elimination procedure based on the log-likelihood ratio test for removal. Once we succeeded in developing a final model, we dichotomized the continuous variables in the model using a cutoff point driven by a univariate classification tree procedure. Again, we checked for refitting the model using these new dichotomized variables. We used both discrimination and calibration to assess this multivariate model’s performance. We assessed discrimination ability by calculating the area under curve (AUC) the receiver operating characteristic (ROC).

The Hosmer–Lemeshow test was used to assess the goodness of fit of the proposed multivariate model. Internal validation of the prognostic model was done using the bootstrap procedure, including random sampling from the source population. In our case, we created 1,000 samples of the same sample characteristics as the study sample size using the replacement method. Then we calculated bias-corrected and accelerated values along with 95% CIs for all the regression coefficients (β). After that, we rounded off the regression coefficients of the final model to derive the final prognostic score, which clinicians can easily use. Then we again evaluated the performance of this bedside prognostic score using the AUC of the ROC curve.

RESULTS

In our study, we included 94 patients (7.6 ± 2.1 years, 69% boys) with TBM to develop and validate the bedside prognostic score. A comparison of baseline sociodemographic, clinical, laboratory, and neuroimaging characteristics of the subgroups with the good and poor functional outcomes are shown in Table 1. Twelve patients died during the follow-up period of 6 months (all during the initial hospital admission and none at home), and 39 patients had poor functional outcomes when assessed at 6 months. The causes of in-hospital deaths in children with TBM were multifactorial. Nosocomial sepsis, secondary respiratory or urinary tract infections, and subsequent septic shock were present in a significant proportion of patients. The duration of intensive care unit stays, associated malnutrition, cross-infection from animate and inanimate sources inside the hospital, rather than primary illness (TBM), were the major causes of mortality.

Table 1.

Univariate analysis showing comparison between patients with poor and good functional outcome (N = 94)

Variable	Poor functional status (N = 39)	Good functional status (N = 55)	P value
Age (years)	5.3 ± 1.5	9.2 ± 2.4	< 0.0001
Gender
Male	29	36	0.37
Female	10	19
Socioeconomic status
Lower	31	39	0.47
Middle	8	16
Higher	0	1
Residence
Rural	35	46	0.56
Urban	4	9
Fever	38	54	0.76
Headache	39	54	0.78
Seizures	18	23	0.83
Refractory seizures*	11	3	0.002
Vision impairment	8	11	0.43
Any focal deficit	22	19	0.056
Cranial nerve palsy (III, IV, or VII)	17	13	0.04
Hemiparesis	9	6	0.15
Paraparesis	2	2	0.11
Papilledema	30	36	0.26
Tonic motor posturing	25	4	< 0.0001
Baseline grade of TBM
I	0	5	< 0.0001
II	1	21
III	38	29
GCS at presentation	9.1 ± 1.6	13.2 ± 1.1	< 0.0001
Baseline PCPC score	4.4 ± 0.7	2.3 ± 0.6	< 0.0001
Meningeal signs	31	43	0.29
Weight/age Z score	−2.1 ± 0.5	−1.6 ± 0.5	0.11
Hydrocephalus	33	39	0.14
Hydrocephalus with periventricular ooze	32	17	< 0.0001
Brain infarction in neuroimaging	17	7	0.0014
Basal exudates	22	29	0.31
Tuberculomas	21	29	0.24
Leptomeningeal enhancement	28	39	0.26
Diagnostic criteria
Definite	8	7	0.62
Probable/possible	31	48
Drug resistance (MDR/R resistance)	1	0	0.81
HIV positivity	1	0	0.81
PCPC score
2	0	9	< 0.0001
3	1	44
4	36	2
5	2	0
CSF cell count (cells/µL)	278.4 ± 37.1	249.5 ± 23.9	0.10
CSF protein (mg/dL)	291.5 ± 31.9	276.1 ± 28.4	0.21
CSF sugar (mg/dL)	38.4 ± 5.7	41.2 ± 6.8	0.58

CSF = cerebrospinal fluid; GCS = Glasgow Coma Scale; MDR/R = multidrug resistant/rifampicin; PCPC = Pediatric Cerebral Performance Category; TBM = tuberculosis meningitis.

^*Requiring more than two antiseizure medications.

Although a large number of variables were included in the univariate analysis, only the following 11 variables at baseline were found to be associated with poor functional status: younger age, stage III TBM, lower GCS at presentation, baseline PCPC score, severe underweight for age at presentation, hydrocephalus with periventricular ooze, presence of brain infarction in neuroimaging, presence of tonic motor posturing, cranial nerve palsy (of cranial nerves III, IV, VI, VII), presence of hemiparesis and papilledema at presentation. However, when these 11 variables were subjected to multivariate binary logistic regression for the development of a prognostic model, younger age, stage III TBM, baseline PCPC score, presence of tonic motor posturing, and presence of brain infarction in neuroimaging (five variables) were independent predictors of poor functional outcome (dependent variable) (Table 2). We dichotomized the continuous variables such as age (< 4 years and ≥ 4 years) using the procedure mentioned in the Methods section. We used a similar method to determine the cutoff for the PCPC score (< 4 and ≥ 4).

Table 2.

Regression coefficients with 95% CI after Bootstrap internal validation for the prognostic model

Bootstrap full model
Variable	β coefficient	Boot strap (BCA 95% CI)	P value
Age < 4 years	0.032	0.017–0.054	0.002
Stage III TBM	1.43	0.61–2.74	0.001
GCS at presentation	−0.38	−0.95 to −0.12	0.02
Baseline PCPC score ≥ 4	1.12	0.48–2.37	0.006
Tonic motor posturing	0.67	0.18–1.43	0.003
Hydrocephalus with periventricular ooze	0.43	−0.37 to 0.79	0.41
Brain infarction in neuroimaging	0.59	0.11–1.78	0.01
Refractory seizures*	0.26	−0.13–1.09	0.38
Cranial nerve palsy	0.16	−0.23–0.91	0.47
Hemiparesis	0.47	−0.38–1.29	0.26
Papilledema	0.06	−0.49–0.81	0.23
Constant	5.23	3.47–8.03
Bootstrap final model
Age < 4 years	0.031	0.015–0.053	0.003
Stage III TBM	1.41	0.59–2.71	0.002
Tonic motor posturing	0.66	0.15–1.41	0.005
Brain infarction	0.57	0.09–1.76	0.02
Baseline PCPC score ≥ 4	1.14	0.47–2.34	0.009
Constant	5.21	3.42–8.01

ASM = antiseizure medications; BCA = bias-corrected and accelerated; CI = confidence interval; PCPC = Pediatric Cerebral Performance Category; TBM = tuberculosis meningitis.

^*Requiring more than two antiseizure medications.

We retested the model using age and PCPC score as dichotomous variables, and all the variables were found to be predictors of poor functional outcomes. The final model incorporating these five variables showed good discrimination in differentiating patients with good and poor functional outcome (AUC = 88.2%; 95% CI: 80.1–91.4%, P < 0.001). The Hosmer–Lemeshow test showed good calibration of the final model (concordance between observed and predicted possibilities; P = 0.53). In the internal validation, the final model held well, as suggested by the narrow bias-corrected and accelerated 95% CIs of all the five predictor variables. All five predictors that were found to have a significant correlation with poor functional outcomes continued to have a significant association, even after the bootstrap procedure.

The beta regression coefficients of all five predictors were multiplied by 3 so that we could round them off quickly to the nearest integers. We then developed a final model containing our bedside prognostic score, as shown in Table 3. We created the acronym “PITAS” to this score we were developing: PCPC score at baseline, infarction on neuroimaging, tonic motor posturing, age, and stage of TBM. The final score was calculated for each patient by adding up the scores assigned to each individual item, determined based on their beta coefficients. The final score ranged from 0 to 12; the higher the score, the greater the probability of a poor functional outcome. We also plotted a nomogram and ROC curve showing sensitivity and 1-specificity at various cut-off points in the score (Figure 1). While the predicted rate of poor functional outcome was 0% for scores 1 through 3, it was 100% for a score of 12. Overall, at baseline bedside score had good discrimination capacity, with an AUC under the ROC of 87.1% (95% CI: 79.9–90.8%, P < 0.01). A score of 10 appeared to be the best cutoff for discriminating between good and poor functional outcomes, with a sensitivity of 81% and specificity of 67% (Table 4). Overall, the score performed better than the MRC stage alone (AUC 82.3%) in predicting functional outcomes (Figure 2).

Table 3.

Bedside prognostic score PITAS from the final model

Variable	Score
Age < 4 years	1
Age ≥ 4 years	0
Stage I/II	0
Stage III	4
Baseline PCPC score 1–3	0
Baseline PCPC score 4–6	3
Tonic motor posturing present	2
Tonic motor posturing absent	0
Brain infarction present on neuroimaging	2
Brain infarction absent on neuroimaging	0

PCPC = Pediatric Cerebral Performance Category; PITAS = PCPC score at baseline, infarction on neuroimaging, tonic motor posturing, age, and stage of tuberculosis meningitis.

Figure 1.

The area under receiver operating characteristic curve shows the performance of the PITAS score at various cut-off points. PITAS = PCPC score at baseline, infarction on neuroimaging, tonic motor posturing, age, and stage of tuberculosis meningitis.

Table 4.

Performance of PITAS score in predicting functional outcome in our cohort of tuberculosis meningitis cases

Total score	No. of patients	No. of patients with poor functional outcome	Predicted probability of poor functional outcome	Sensitivity	Specificity
0–1	5	0	0%	–	–
2	4	0	0%	–	–
3	6	0	0%	–	–
4	10	2	20%	49%	100%
5	12	3	25%	53%	92%
6	12	4	33%	59%	87%
7	11	6	54%	66%	82%
8	10	6	60%	70%	75%
9	8	5	62%	75%	70%
10	7	5	71%	81%	67%
11	5	4	80%	88%	64%
12	4	4	100%	100%	62%
Total	94	39	41%

PITAS = PCPC score at baseline, infarction on neuroimaging, tonic motor posturing, age, and stage of tuberculosis meningitis.

Figure 2.

The area under receiver operating characteristic curve compares the performance of PITAS score and UK Medical Research Council staging in our cohort of pediatric tuberculosis meningitis. PITAS = PCPC score at baseline, infarction on neuroimaging, tonic motor posturing, age, and stage of tuberculosis meningitis.

DISCUSSION

We have developed a simple bedside score that can be used in children with TBM at baseline to predict the functional outcome at 6 months, with acceptable sensitivity and specificity. This PITAS score fares better than the MRC staging system in predicting the functional outcome.

Cerebral infarction was an independent predictor of poor outcomes in TBM. Although more widely available, computed tomography scan can miss small infarcts in the acute stage, and hence MRI is the ideal option for diagnosis.¹⁷ Relying solely on clinical findings may create errors due to other differential diagnoses, such as hemiparesis or ocular motor palsies secondary to brain herniation, tuberculoma, or TB abscess,¹⁸ and this is probably why clinical focal deficit was not found to be an independent predictor in our scoring system. Subjective or objective severity of hemiparesis also often varies and is likely to introduce bias in the scoring system and reduce its accuracy.

GCS score at presentation is a strong predictor of the final outcome, but because it was part of the MRC staging we used,¹⁹ it was not included as another parameter in our scale.

The MASH-P [baseline Modified Barthel index (M), age (A), stage (S), hydrocephalus (H), and papilledema (P)] score was devised by Rizvi et al.¹⁰ for predicting death in adults with TBM. We have avoided using mortality as a key outcome measure because the number of deaths was so small, and the sample size of our cohort was not large enough to develop a prediction score for mortality. Also, all mortalities in our patients occurred in-hospital, and the cause of death was often multifactorial and unrelated to the primary illness (e.g., nosocomial infection, septic shock) with the delay in referral often being a main contributing factor in resource-constrained settings such as ours. We can probably predict mortality in TBM cases by using other validated scores used in critically ill children, such as the Acute Physiology and Chronic Health Evaluation score or Pediatric Index of Mortality, Pediatric Risk of Mortality score. This needs to be explored in future studies.^20,21

We found tonic motor posturing an independent predictor of poor outcome in TBM. We used tonic motor posturing as a surrogate marker of clinically significant raised ICP because most resource-constrained centers do not routinely use ICP catheters.²² Papilledema, unequal pupils, the Cushing triad, and focal deficits are not always seen in patients with raised ICP, whereas we observed that most patients with tonic motor posturing had decompensated hydrocephalus in neuroimaging. Moreover, even with mild hydrocephalus, an acute increase in ICP can occur, but papilledema often requires a few days to develop. Raised ICP can also occur due to the space-occupying effect of tuberculomas, cerebral edema, venous thrombosis, or arterial infarcts.

Although hydrocephalus with periventricular ooze or transependymal edema was one of the predictors in univariate analysis, it did not reach the level of significance in multivariate analysis. This may have been due to the lack of an objective grading system for periventricular ooze. Although periventricular ooze often suggests decompensated hydrocephalus under pressure, this might not be universally true. Sometimes even mild hydrocephalus can be associated with significant raised intracranial pressure, papilledema, and brain herniation. More often, TBM patients have mild hydrocephalus without significant clinical implications. Furthermore, a subset of cases develops an increase in hydrocephalus a few weeks after presentation.

Our study’s prognostic factors for poor functional outcomes are consistent with the previous literature on children and adults in this regard. The study by Karande et al.¹² included 123 children with TBM and had a high mortality rate of 23%. The lower mortality rate in our cases may reflect the improvement in clinical practice over the 17 years between their study and ours.

Delage et al.,⁴ in an older study from almost half a century ago, reported a mortality rate of 38% and poor prognostic factors to be Stage III disease (severe changes in sensorium and often severe neurologic abnormalities) at the time of admission, age ≤ 3, associated miliary tuberculosis, and delay in the initiation of therapy. In contrast, the survey by Mahadevan et al.¹¹ suggested a mortality rate of 10%, which is more in line with our research, and the following poor prognostic factors: younger age, tonic posturing, papilledema, focal neurological deficit, and advanced stage at presentation. In the study by Faella et al.²³ from Italy, 13% of patients died, and 19% patients survived with sequelae. Long-lasting pre-admission nonspecific symptoms, high CSF protein, MRC Stage III, and ventricular dilation were associated with higher morbidity/mortality rates. Faster normalization of CSF parameters was associated with better clinical outcomes. Wang et al.²⁴ found that coma and CSF protein > 1,188.3 mg/L were associated with poor functional outcomes.

Rizvi et al.¹⁰ in an adult study from northern India suggested a mortality rate of 17% and the following as predictors of mortality: higher age, stage III disease, baseline Modified Barthel Index ≤ 12, papilledema, and hydrocephalus. The Haydarpasa Meningitis Severity Index is also designed to prognosticate adult TBM patients.²⁵ However, this scale used many parameters that are not relevant for the pediatric population, such as the MBI and old age in MASH-P score. Moreover, it had several poorly defined parameters such as immunosuppression and vasculitis. None of these scores have been explored or adapted for the pediatric population. However, that study also showed that MRC Stage III staging alone could not predict mortality.

Thao et al.⁹ developed another prognostic model for adult TBM patients, with separate models for HIV-negative and HIV-positive patients. The parameters included in the model for HIV-negative patients resemble the score derived in our study. However, the calculation of this score is complex, requires nomograms, and is unsuitable for the pediatric population.

There are several limitations to our study. We did not have a prespecified sample size because we initially started with clinical variables and did not propose an anticipatory prognostic model. The PCPC score at baseline, as such, has good sensitivity and specificity in discriminating between those with and without poor functional outcomes. However, we intended to develop a prognostic score specific to TB meningitis. We wanted specifically to include prognostic factors related to TB meningitis, which can help clinicians predict poor functional outcomes. Moreover, the PCPC score alone has excellent discriminating ability because the same score was used on the follow-up to differentiate between those with good and poor functional outcomes. The studies in adults that derived various prognostic models used larger sample sizes, but pediatric studies often had a much smaller sample size than ours. Many patients were lost to follow-up, and we could not include them in our prognostic models. CB-NAAT positivity rate and isolation of Mycobacterium were extremely low in our study. Only one patient was HIV-positive, and almost all patients in our study were immunocompetent without any preexisting chronic systemic illness. Thus, the score derived in this study is mainly meant for rifampicin-sensitive TBM cases in immunocompetent children.

Many patients who had poor functional outcomes might be harboring drug-resistant bacilli. Although we have only completed internal validation of our suggested model, external validation by other investigators from different parts of the world is necessary before recommending its use in widespread clinical practice. Still, the PITAS score developed and internally validated in the current study is a novel attempt at developing a prognostic scoring for pediatric TBM cases.

CONCLUSION

The PITAS score, which was internally developed and validated, has good sensitivity and specificity in predicting poor functional outcomes in pediatric TBM cases at 6 months. However, it needs to be externally validated in different populations.