Can improved diagnostics reduce mortality from Tuberculous meningitis? Findings from a 6.5-year cohort in Uganda [version 1; peer review: 3 approved with reservations]

Background: Tuberculous meningitis (TBM) is the second most common cause of meningitis in sub-Saharan Africa and is notoriously difficult to diagnose. We describe the impact of improved TBM diagnostics over 6.5 years at two Ugandan referral hospitals. Methods: Cohort one received cerebrospinal fluid (CSF) smear microscopy only (2010-2013). Cohort two received smear microscopy and Xpert MTB/Rif (Xpert) on 1ml unprocessed CSF at physician discretion (2011-2013). Cohort three received smear microscopy, routine liquid-media culture and Xpert on large volume centrifuged CSF (2013-2017) for all meningitis suspects with a negative CSF cryptococcal antigen. We compared rates of microbiologically confirmed TBM and hospital outcomes over time. Results: 1672 HIV-infected adults presenting with suspected meningitis underwent lumbar puncture, of which 33% (558/1672) had negative CSF cryptococcal antigen and 12% (195/1672) were treated for TB meningitis. Over the study period, microbiological confirmation of TBM increased from 3% to 41% (P<0.01) and there was a decline in Open Peer Review


Introduction
Tuberculous meningitis (TBM) is the second most common cause of adult meningitis in sub-Saharan Africa 1,2 , accounting for one to five percent of the 10.4 million tuberculosis (TB) cases reported worldwide in 2016 3 . Despite treatment, TBM outcomes are poor with 19-28% mortality in HIV-uninfected persons and 40-67% mortality in HIV-infected patients in addition to long-term disability is frequent among survivors 4-6 .
Insidious symptom onset in persons with TBM leads to delay in seeking care and increasing disease severity at presentation correlates with higher mortality 7 . Further, the paucibacillary nature of TBM increases the difficulty in confirming diagnosis once care is sought, also contributing to high mortality 8 . Cerebrospinal fluid (CSF) smear microscopy for acid-fast bacilli (AFB smear) has poor sensitivity (~10-20%) in routine practice 7 . Culture has improved sensitivity (~50-60%) but is not widely available in many resource constrained settings and commonly takes at least 2-3 weeks for liquid culture growth, which is too slow to guide decision-making at the time of presentation 8 .
In 2013, the World Health Organization endorsed the Xpert MTB/RIF (Xpert) assay (Cepheid, Sunnyvale, California, USA), a cartridge-based, polymerase chain reaction assay with a run time of 113 minutes, as the preferred initial test to investigate TB meningitis on the basis of a meta-analysis of 13 studies 9 . Of the two major studies included in the meta-analysis, Patel and colleagues reported 67% sensitivity against microbiologically proven TBM and 36% against consensus clinical case definitions, while Nhu and colleagues showed 59% sensitivity against the same case definitions 10-12 . Additionally, use of a larger volume of centrifuged CSF improves sensitivity of Xpert 10,13 . Yet, inadequate negative predictive value means that Xpert cannot substitute for clinical judgement 14 .
There is evidence that use of Xpert for diagnosis of pulmonary TB reduces diagnostic delay, increases the rate of same day treatment, and decreases usage of empiric treatment 15,16 . However for pulmonary TB, Xpert has not been shown to decrease mortality [16][17][18] . Yet, lessening diagnostic delay in persons with TBM may be more likely to lead to improved outcomes as compared to pulmonary TB given the high early mortality of TBM 19 . Whether routine use of Xpert for investigation of suspected TBM has made an impact on mortality has not yet been investigated.
Herein we describe TBM diagnosis and outcomes over a 6.5-year period in prospective cohorts at two Ugandan referral hospitals.

Study population
Adults presenting with symptoms of meningitis to Mulago National Referral Hospital, Kampala, and Mbarara Regional Referral Hospital, Mbarara, Uganda were assessed for eligibility for enrolment in two consecutive clinical trials investigating cryptococcal meningitis (www.clinicaltrials.gov: NCT01075152 -Cryptococcal Optimal Antiretroviral Timing, NCT01802385 -Adjunctive Sertraline for the Treatment of Cryptococcal Meningitis) beginning November 22, 2010. Patient screening continued until 28 th May 2017. Baseline demographics, clinical and outcome data on all screened patients were routinely collected as part of the screening process. Any patient who received testing for TBM was eligible to be included in this study.
Microbiologically proven (definite) TB meningitis was defined as any positive AFB smear, culture or Xpert result from CSF testing. Consensus uniform case definitions were used to categorise patients as definite, probable, possible or not TBM 11 . In addition to the consensus case definitions, clinical TBM was defined as subjects without a positive microbiologic result, but who were treated empirically for TBM treatment due to high clinical suspicion. TBM treatment included 12 months of antituberculous therapy with 6-8 weeks of adjunctive corticosteroids as per Ugandan guidelines 20 .
Cohort definitions and diagnostic tests used Cohort one (16th November 2010 until 28 th May 2013) received only CSF AFB smear testing ( Figure 1). If available, 1mL cryopreserved CSF was later tested with Xpert MTB/Rif when Xpert became available. Cohort two (1 st April 2011 until 10 th November 2013) underwent CSF AFB smear and Xpert MTB/Rif on a 1ml sample of uncentrifuged CSF. Testing was performed at physician discretion when there was lymphocytic pleocytosis and/or high degree of clinical suspicion. In the period of overlap of cohort one and two (April 2011-May 2013), Xpert testing was not being done on a routine basis; subjects were included in cohort two when Xpert was done in real-time and in cohort one if Xpert was not done, or only done at a later date on cryopreserved specimens.
In cohort three (11 th November 2013 until 28 th May 2017) all cryptococcal antigen negative (IMMY, Norman, Oklahoma, USA) patients were systematically investigated for the presence of TB meningitis, irrespective of physician discretion. Subjects had comprehensive testing for TBM with CSF AFB smear (Mulago Hospital only), Xpert MTB/Rif on large volume centrifuged CSF 13 and CSF Mycobacteria Growth Inhibitor Tube culture (MGIT, Becton Dickinson, Franklin Lakes, USA). AFB smear was discontinued in Mbarara in 2013 as the sensitivity was deemed too low to justify further use. In patients with a confirmed diagnosis of cryptococcal meningitis (CM), if TBM co-infection was suspected, patients would be investigated for TBM at the physician's discretion.

Assessment of outcome
In-hospital outcome was determined from case report forms, hospital medical records or follow-up telephone calls with the patient or their surrogate where hospital outcome was unknown. The outcome was categorised as discharged alive, deceased prior to hospital discharge or unknown (i.e. self-discharged against medical advice in an imminently terminal patient, hospital outcome undetermined, transferred to another facility).

Statistical methods
Comparisons of categorical and continuous demographic and clinical characteristics by cohort were performed using Fisher's exact tests and Kruskal-Wallis tests, respectively. Sensitivity of Xpert MTB/Rif was evaluated against a composite reference standard (any positive CSF test -AFB smear, Xpert or culture i.e. definite TBM according to the uniform case definition) 11 . A separate analysis was conducted against the uniform case definition of probable or definite TBM 11 . Concordance between Xpert MTB/Rif and culture was evaluated with a kappa statistic and McNemar's test. Invalid tests (e.g. culture contamination, Xpert error) were counted as negative results. Mortality was first compared by cohort for participants with a known outcome using Fisher's exact test. Data for patients with unknown outcome was imputed to assume first that 50% within each cohort died, or that 75% died (both within the expected mortality range for this population). Odds ratios and 95% confidence intervals were computed from multivariable logistic regression models with these imputed data, adjusted for 1) ART status, and 2) ART status and definite TBM diagnosis. Imputations were repeated with new random assignments to confirm results. Analyses were conducted using SAS version 9.4 (The SAS Institute, Cary, NC) and p-values <0.05 were considered statistically significant. , and by the Uganda National Council of Science and Technology Written informed consent for screening or participation in the studies was obtained from all participants or from their surrogates (e.g. family member or guardian) where the patient had altered mental status and did not have the capacity to provide consent.

Participant characteristics
Over the study period, 1672 patients with meningitis symptoms were assessed and underwent lumbar puncture: 1058 (63%) had a positive CSF cryptococcal antigen test, 558 (33%) had negative CSF cryptococcal antigen test (data missing, n=56). A total of 195 subjects were treated for TBM. Overall 61% were male, median age was 35 years (IQR 30-42), 96% were HIV-positive, median CD4 count was 78 cells/μL  and the majority (69%) presented with British Medical Research Council severity grade II disease, see Table 1. Baseline characteristics were similar between cohorts with the exception of antiretroviral (ART) experience; 0% of participants were on ART in cohort one compared to 61% in cohort three (P<0.01).

Outcomes
Hospital outcome was known for 142 participants, 53 had unknown outcomes or self-discharged against medical advice.
Median time to death was 3 days (IQR 1-9 days) among those known to have died, and median length of hospitalization was 7 days (IQR 4-10 days) for participants known to have survived to hospital discharge. Among those with known outcomes, there was a non-significant decline in mortality from 57% in cohort one to 41% in cohort three (p=0.27) (  Figure 3).

Discussion
Rapid molecular diagnostics have been predicted to reduce TB-related mortality 21 but no prior studies have looked at the impact of Xpert on TBM-related mortality. Here we report clinical outcomes among hospitalized Ugandans treated for TB meningitis over a 6.5-year period. In-hospital mortality   2.0 (0.6,6.7) 1 Overall median (IQR) time in hospital was 7 (4, 10) days among those who were known to be discharged alive, and 3 (1, 9) days among those who were known to have died in hospital *P-value from Fisher's exact test comparing KNOWN discharged alive vs KNOWN died; Odds ratios are the odds of being discharged alive, assuming 50% and 75% of those with unknown outcome died was high in the cohort overall (44%), similar to other research settings with high HIV prevalence 7,19,22,23 . The adjusted model on imputed data found odds of in-hospital mortality were almost two fold higher in the earliest cohort, tested by CSF smear microscopy only, than that of that of the most recent cohort in whom Xpert (and culture) are routinely performed. Severity of TBM at presentation was similar over the study period and TBM treatment recommendations have not changed for Uganda, though other confounding factors may exist.
Although we lack data regarding time to treatment initiation, in this research setting Xpert results were obtained within 24 hours, leading to prompt treatment in the 51 subjects who were positive by Xpert MTB/Rif in real-time and presumably contributing to reduced mortality. The proportion of ART experienced subjects increased significantly over time with the roll out of ART treatment in Uganda and because the parent trial in cohort one enrolled only ART naïve subjects 24 . Although ART status was not associated with mortality, we did adjust for ART in multivariable models due to the large discrepancy in ART status between cohorts.
Despite a non-significant decline in mortality, a current casefatality rate of 41% remains unacceptably high and highlights the remaining work required to achieve the WHO goal of reducing TB-related deaths by 90% by 2030 25 . Initiating treatment in the early stage of disease is the single most important factor in improving outcomes 7 . Earlier presentation to the hospital is essential for prompt diagnosis and treatment initiation, yet, 83% of our cohort presented with MRC grade II or III disease.
Once the patient presents to care, an affordable, rapid, and reliable test that can effectively confirm or rule out TBM is crucial for prompt diagnosis. In this predominantly HIV-positive TBM cohort, sensitivity of Xpert was 63% against the composite reference standard. Thus, even though results were available rapidly, Xpert missed over one in three cases. The next generation assay Xpert MTB/Rif Ultra has an analytic limit of detection of 15 colony forming units (CFU)/ml, compared to 113 CFU/ml for Xpert 26 . Ultra appears to be significantly more sensitive than Xpert or culture for the diagnosis of TBM (95% versus 45% and 45% respectively, P<0.001) 27 . Whether Ultra can reduce diagnostic delay and improve outcome from TBM requires further prospective evaluation.
Where both Xpert and MGIT had been done, less than a third (23%, 17/74) of confirmed cases were positive by both modalities. This is consistent with prior findings and is likely due to the relatively higher sensitivity of culture versus Xpert, and the ability of Xpert to detect dead TB bacilli 13,27 . Neither test performed better than the other (P=0.42). . Illustration of odds of dying in cohort one and two compared to cohort three in a multivariate model. Odds ratios (and 95% confidence intervals) for death by the end of hospitalization comparing cohorts 1 and 2 to cohort 3, computed from multivariable logistic regression models with imputed data, adjusted for (1) ART status, and (2) ART status and definite TBM diagnosis. Data for patients with unknown outcome was imputed to assume that 50% within each cohort died, or that 75% died.
Until a highly sensitive assay is widely available, there is likely to be on-going heterogeneity in clinical practice regarding initiation of empiric therapy for TBM. In our study, Mulago Hospital participants were treated for TBM on an empiric basis in under one quarter of cases as opposed to over three quarters of cases at Mbarara Hospital. Though empiric TBM therapy is potentially life-saving, significant risks such as side effects, drug-interactions and adjunctive steroids in an already immunosuppressed population need to be considered. Ideally, a rapid, accurate test allows therapy for TB meningitis to be started promptly only in those who actually have TBM. Overall, the proportion with microbiologically confirmed TBM increased significantly from 3% in cohort one to 41% in cohort three (P<0.01). In cohort two, Xpert was only performed in cases where there was extremely high index of suspicion and empiric treatment was given only twice in those with a negative Xpert (4%, 2/56). The low number of empiric diagnoses during this period were likely due to over-confidence in Xpert's ability to rule-out TBM. As understanding regarding the limitations of Xpert for the diagnosis of TBM became known, empiric TBM treatment rose 14 .
Limitations of this study include missing data on hospital outcomes and time to starting TB treatment, unbalanced numbers in each cohort including smaller numbers in earlier cohorts and lack of long-term outcome data. When imputing data in the model we assumed that either 50% or 75% of patients with unknown outcome actually died, which is a clinically reasonable judgment for this population 4 .
Here we present important data on rates of diagnostic confirmation and TBM mortality during a period of TB diagnostic evolution. There has been a significant increase in microbiological confirmation and a modest, albeit non-significant, decline in mortality since introduction of Xpert and culture in our study setting. An on-going multifaceted approach is needed to further reduce death and disability from TBM.

Consent
Written informed consent for publication of the anonymised data was obtained from the participants or their surrogates.

Data availability
The database contains individual level data and as such is not available through an open-access data repository. The database is stored on a secure server at University of Minnesota. Researchers interested in accessing the data can contact the corresponding author (FVC), the last author (DRB) or the Division of Biostatistics at the University of Minnesota. Data access will be granted to active researchers in the field with the agreement of the authors.

Competing interests
No competing interests were disclosed. Many thanks for requesting us to review this article which addresses an important unanswered question in TB diagnostics. We have made specific points regarding each section of the manuscript as well as giving an overall impression.

Grant information
This prospective cohort studies observes differences in diagnosis and hospital outcomes related to TBM across three cohorts of patients recruited over a 6.5 year period in Uganda. The study aims to measure the impact on diagnosis and hospital outcomes with the addition of Gene Xpert testing or Gene Xpert androutine liquid media culture to standard diagnostic testing (CSF smears) over time.

Abstract
The study design is unclear: is this a retrospective or post-hoc analysis of data collected from other studies? The nature of the parent studies should be mentioned.

○
The denominator is not clearly reported: the inclusion criteria for the current analysis appears to be "all meningitis suspects with a negative CSF cryptococcal antigen." Were 558 patients with negative CSF cryptococcal antigen test included in the analysis? ○ The conclusion that "in-hospital mortality has declined" seems exaggerated because this trend towards lower mortality over the study periods was not significant, and may have been due to other factors unrelated to changing diagnostics.

Methods
The inclusion criteria should be clearly defined and justified.
○ Differences between the cohorts could have introduced bias towards improved outcomes (type 1 error): cohorts 1 and 2 at physician discretion, cohort 3 were systematically investigated. This should be addressed in the discussion.

Results
The denominator is not clearly reported: the inclusion criteria for the current analysis appears to be "all meningitis suspects with a negative CSF cryptococcal antigen." Were 558 patients with negative CSF cryptococcal antigen test included in the analysis, or was it 195 subjects that were treated for TBM? This should match the inclusion criteria, which appear to be any patient with suspected meningitis and negative CrAg testing. The paragraph describing empiric treatment practices requires clarification: how was 'empiric treatment' defined? The numbers appear to suggest that no patients with a confirmed diagnosis were empirically treated (ie empiric treatment = no microbiological confirmation). Suggest changing the terminology to make this clearer -1. Treated based on positive result; 2. Treated without a positive result. It seems that in Mulago Hospital a higher proportion of treated patients had positive results, and therefore did not require 'empiric treatment.' So, the important question seems to be why there was such a bug discrepancy in the number of confirmed cases between the two sites, allowing 'empiric treatment' to be avoided?

○
The denominator is again unclear for the diagnostic accuracy section: where do the "455 tested" come from if only 195 cases were included in the study (or is it 558?). Suggest including a consort diagram to explain the patient populations for this study. ○ Suggest including a 2 x 2 table to demonstrate the performance of diagnostic tests. Although the sensitivity of Xpert is 63% compared to a reference standard, this this standard is only present in 40% of cases (60/147) in Cohort 3.

Discussion
See comments related to discussion above ○ The main (and probably only data-driven) conclusion is that more confirmed TBM cases were detected with the use of Xpert and culture. The authors did not convincingly show an improvement in outcomes. Any observed trends could be secular, or related to other confounding factors resulting from different populations, selection criteria, management practices, etc.

○
The study observes three cohorts recruited from two parent trials. Although the data provides an important insight into the unanswered questions above, the suitability of the cohorts as comparators to one another does pose some limitation to how the results can be interpreted. Namely there is large variation between the observed cohorts in a number of aspects namely the size of cohorts, ART status, and availability of follow up data. Although ART status had been accounted for in analysis, the impact of absent follow up data which is much more prominent for cohort 1 (79%) compared to cohort 3 (16%) should be made clearer in the discussion. Also, how were the patients allocated to cohort 1 and 2 during the overlapping study period (1 st April 2011 until 28 th May 2013)? Were those where there was more diagnostic doubt allocated to receive testing with GeneXpert (and therefore subsequently included in cohort 2 than cohort 1?).. could this have affected outcomes observed between the two groups?
It is possible that apart from the mentioned variables, that over time, other aspects may have influenced hospital outcome in these patients over a period where there has naturally been some advances made in TBM care besides the diagnostics discussed. For example the use of imaging diagnostics, time to treatment, availability of drugs, better understanding of supportive measures in TBM. I feel that these should form part of the discussion here and although they may not be quantifiable in this setting, should be acknowledged as possible factors for the observed differences over time. In this setting specifically were there changes in resource allocation? Did hospital facilities change?

If applicable, is the statistical analysis and its interpretation appropriate? Yes
Are all the source data underlying the results available to ensure full reproducibility?
No source data required

Are the conclusions drawn adequately supported by the results? Partly
Competing Interests: No competing interests were disclosed.
We confirm that we have read this submission and believe that we have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however we have significant reservations, as outlined above. The nature of the parent studies has been expanded in the methods section of the revised manuscript. We have also added further information to the abstract and clarified that this is a post-hoc analysis of prospective cohorts.

The denominator is not clearly reported: the inclusion criteria for the current analysis appears to be "all meningitis suspects with a negative CSF cryptococcal antigen." Were 558 patients with negative CSF cryptococcal antigen test included in the analysis?
The main analysis for microbiological confirmation and mortality relates to the 195 people within the three TBM prospective cohorts. Of the 1672 screened any patient who received any testing for TBM (CSF AFB smear, Xpert or mycobacterial culture) during this period was eligible to be included in table 2 and the diagnostic accuracy of Xpert and culture analysis. Any patient who was ultimately treated for TBM (n=195) was eligible to be included in the three prospective TBM cohorts, and it was this group on whom the rate of microbiological confirmation and outcomes were compared. We have added a schematic diagram to the manuscript which i hope clarifies the nature of the study population.

The conclusion that "in-hospital mortality has declined" seems exaggerated because this trend towards lower mortality over the study periods was not significant, and may have been due to other factors unrelated to changing diagnostics
You are absolutely correct. We have changed to text in the abstract to read "Since 2010, as TBM diagnostics have evolved, microbiologically-confirmed TBM diagnoses have increased and there has been a non-significant decline in TBM in-hospital mortality. Due to multiple possible confounding factors it is not possible to conclude what has driven this decline in mortality".

The inclusion criteria should be clearly defined and justified
The following text has been added to the revised manuscript: "Adults presenting with suspected meningitis (headache and neck stiffness +/-vomiting, fever, seizures, focal neurological deficits, or altered consciousness), to Mulago National Referral Hospital, Kampala, and Mbarara Regional Referral Hospital, were assessed" "Participants with non-cryptococcal meningitis were not enrolled into the clinical trials but followed until hospital discharge". "Any patient who was ultimately treated for TBM was eligible to be included in one of the three TBM cohorts, on which rates of microbiological confirmation and outcomes were compared. Cohort was determined by what type of TB testing they had undergone".

Differences between the cohorts could have introduced bias
The review by Dr Hamers also made this comment. Please see response to his comment for a potential explanation.
5. The paragraph describing empiric treatment practices requires clarification: how was 'empiric treatment' defined? The numbers appear to suggest that no patients with a confirmed diagnosis were empirically treated (ie empiric treatment = no microbiological confirmation). Suggest changing the terminology to make this clearer.
Thank you. We have defined this more clearly in the methods.
6. So, the important question seems to be why there was such a big discrepancy in the number of confirmed cases between the two sites, allowing 'empiric treatment' to be avoided?
Both hospitals have the same available diagnostics so i believe the difference in empiric treatment relates to local or personal thresholds applied to the initiation of TB treatment in the absence of a confirmatory test. You will no doubt have observed different treatment thresholds amongst colleagues, especially faced with a critically ill patient. As Mbarara Hospital is a relatively small hospital with good longterm retention of staff, especially in the clinical research setting, it is possible that a handful of clinicians may have seen the majority of TBM cases and been more willing to initiate TB treatment despite negative tests than in Kampala where there is a bigger team and clinicians rotate regularly.

Suggest including a consort diagram to explain the patient populations for this study.
Thanks for this suggestion. We have included a schematic of patient flow from the screening population into the cohort. It is not a classic consort diagram since this population were not part of an RCT.

Suggest including a 2 x 2 table to demonstrate the performance of diagnostic tests.
We have added a 2x2 table.

The main (and probably only data-driven) conclusion is that more confirmed TBM
cases were detected with the use of Xpert and culture. The authors did not convincingly show an improvement in outcomes. Any observed trends could be secular, or related to other confounding factors resulting from different populations, selection criteria, management practices, etc.
Indeed. We acknowledge the limitations of the data available to us and have tempered the conclusion accordingly in the revised manuscript.
2. For example the use of imaging diagnostics, time to treatment, availability of drugs, better understanding of supportive measures in TBM. I feel that these should form part of the discussion here and although they may not be quantifiable in this setting, should be acknowledged as possible factors for the observed differences over time. In this setting specifically were there changes in resource allocation? Did hospital facilities change?
This a good thoughts and we concede that many potential confounders exist. However, we could show from our limited data that severity at presentation across cohorts was the same. Whilst the health services in the private sector have evolved in this time period the government facilities remain under resourced without access to routine blood tests and imaging. Thankfully access to TB medication has been stable for many years. Healthcare worker practices may certainly have changed and we recognise this as a potential confounder.
Thank you again for your detailed review. I hope the responses and amendments are satisfactory.
historical cohorts that span the time period of 2011-2017, and I feel the authors could present these complex factors in a more structured and comprehensive manner in the Discussion: -For instance, the main driver of the difference observed is the huge discrepancy between ART status across cohorts (0-61% on-ART) -is it at all possible to adjust for this major difference with statistical methods alone? -The authors note no major changes in treatment practices over time, although they also report substantial differences between the 2 hospitals in ability to confirm diagnosis and start empirical treatment, which may have impacted on outcomes (if varied with time).
-Referral practices and general awareness among care providers may have improved in the study setting (the authors note they do not have data on time to Rx initiation), especially given that the authors have implemented 2 large clinical trials in the same period.
-The patient case mix may have changed over time in many other ways (e.g. better nutrition, better education, lower TB incidence, better HIV diagnosis and treatment, etc) My recommendation is to reword the title, to avoid misleading the readership.
2. The authors have (deliberately) done separate analyses on the 3 cohorts, and combining historical cohorts may have its limitations. Nonetheless, the combined cohort offers opportunities to attempt answering the main question. In this respect, it would be interesting if the authors could undertake a multivariate analysis of the combined cohort data to identify what are the main determinants of mortality (adjusting for time period). This could help to establish whether indeed the diagnostics influenced outcomes, or that the survival gain was mainly driven by other factors (e.g. ART). In my view, adding this analysis could strengthen the paper.
Minor points 1. The % of microbiological confirmation is 3% in cohort #1, 87% in cohort #2, and 41% in cohort #3. It would be relevant to learn why the latest % seems to be lower than the middle one (despite better diagnostic protocols).
2. The degree of agreement between Xpert and MGIT is very low (kappa 0.23). I would like to see a better explanation for this discrepancy, and how this knowledge can to be applied in recommendations for testing algorithms in practice.

If applicable, is the statistical analysis and its interpretation appropriate? Yes
Are all the source data underlying the results available to ensure full reproducibility?
primary variable of interest for this study. Since ART use changed over time and definite TBM is associated with increased risk of mortality, these covariates were chosen to be included in adjusted analyses. Mortality was unknown for 27% of participants, so we conducted a series of models with imputations assuming that 50% within each cohort died, or that 75% died (both within the expected mortality range for this population). Six multivariable models of in-hospital mortality were run and all cohorts were included in all models. Cohort 3 was chosen as the reference category for presenting the odds ratios of interest since it represents the most current testing era. In all models, neither ART nor definite TBM status had a significant association with in-hospital mortality. All models run include all cohorts and the odds ratios comparing cohorts from the adjusted models are presented in Figure 4 in the (revised) manuscript.
4. The % of microbiological confirmation is 3% in cohort #1, 87% in cohort #2, and 41% in cohort #3. It would be relevant to learn why the latest % seems to be lower than the middle one (despite better diagnostic protocols).
This is an interesting observation and we believe relates to two factors: a) In cohort 2, during the initial period after Xpert introduction, it was only being used on cases with an extremely high clinical index of suspicion so there was a selection bias in population tested.
b) The % of confirmed case amongst the total number treated for TBM in cohort 2 appears spuriously high as the amount of empiric treatment was very low during the cohort. One potential reason is that there was an over confidence in Xpert's ability to be able to rule out TBM. In subsequent years a number of papers showed that the negative predictive value of Xpert is such that a negative test result should not deter TB treatment and empiric treatment was used more frequently again in out setting.

The degree of agreement between Xpert and MGIT is very low (kappa 0.23). I would like to see a better explanation for this discrepancy, and how this knowledge can to be applied in recommendations for testing algorithms in practice.
This is another interesting observation, and a somewhat surprising finding to us also. We have not elaborated in the discussion in the interest of brevity but possible explanations are mentioned here: Positive on Xpert whilst negative on culture: We lack data about the number of cases in the cohort who were on TB treatment at the time of lumbar puncture. It is plausible that if TB treatment had been initiated in the days or weeks prior to the CSF analysis the mycobacterium may have been rendered non-viable for culture. In our setting many patients present with a disseminated TB picture, some of whom have been coughing for several weeks prior to development of neurological symptoms and already initiated on antituberculous therapy from outpatient settings. We hope to capture this data more comprehensively in prospective studies. Positive on culture whilst negative on Xpert: Culture has a lower limit of detection that Xpert so patients with bacillary burdens on the <100CFU/ml region would likely be identified by culture only.

Considerations for future testing algorithms:
It is hopeful that the next generation assay Xpert MTB/Rif 'Ultra' will have a limit of detection similar to that of culture, so given that it can also detect non-viable bacilli, it is likely to perform better the culture in the diagnosis of TBM (as was the case in Bahr N. Lancet ID. 2018). Testing algorithms must be customised and take into consideration the setting, HIVprevalence, cost and the volume of CSF available for testing, amongst other things. Splitting a small CSF sample between multiple assays is likely to be counterproductive. In our setting we would recommend Ultra as the initial CSF test for TBM (in patients with a negative CSF cryptococcal antigen lateral flow assay). Culture remains an important adjunctive test especially in patients where there is a risk of drug resistance.
We have submitted a revised manuscript and hope these changes and responses are to Dr Hamer's satisfaction.

Tom Boyles
Wits Reproductive Health and HIV Institute, University of the Witwatersrand, Johannesburg, South Africa Thank you for the opportunity to review this article which describes the investigation, treatment and outcomes of a cohort of mostly HIV infected adults in Uganda who were suspected of having TB meningitis. It is well written and easily reproducible by others.
The title asks the question of whether improved diagnostics can reduce mortality in TBM. It might be better to say 'Have improved diagnostics improved outcomes?' My view of the data is that it provides very limited evidence that improved diagnostics have improved outcomes of TBM in this cohort. The small improvements in outcome over time are statistically non-significant and due to the observational design are likely to be influenced by numerous unmeasured confounding factors. For example, the experience of clinician may well have changed over time, we know that empiric treatment was used differently in the 2 centres but it might also have changed over time.
Part of the work-up of a patient with suspected TBM is a search for extra-neural TB and we do not know how this changed over time although it is likely that Xpert MTB/RIF was used for non-CSF samples in later cohorts compared to earlier, the same might be said for urine LAM.
My view as a reader is therefore that this data does not provide convincing evidence that the change in diagnostics was the driver of the small changes in mortality. It must be remembered that new diagnostics also have the potential to worsen outcomes, particularly if clinicians miss-interpret negative tests as ruling out the condition as might well occur with Xpert MTB/RIF and TBM.
I think that this well written work definitely deserves to be published but feel that the conclusions should reflect a greater level of uncertainty.
Minor points: The abstract says that all 1672 patients were HIV infected but in the results section is says 96% were HIV infected so there is a discrepancy.
Last line, para 1 of introduction, probably remove word 'to'.
Last line. Para 3 of introduction-It is not that Xpert cannot substitute for clinical judgement-Xpert is used to enhance clinical judgement, it's just that the poor sensitivity means that when negative it has limited influence on decision making Statistical methods-the uniform case definition is probable or definite TBM-do the authors think this is a reasonable reference standard for other tests? My view is that the threshold for treating TBM is very low and that patients with possible/probably or definite TBM should receive treatment and therefore this would be a more appropriate reference standard-what are your views on that?
Methods-Not clear if there were 76 or 74 microbiologically proven cases Discussion-Para 1, repeat of 'that of' Discussion-Para 4, First sentence implies that tests are the only answer to the diagnosis of TBM, what about clinical prediction rules relying on clinical data-although so far they have not been very successful, neither have tests so a robust CPR may negate the need for tests.
Discussion-Para5, Do the authors really think there are dead bacilli in CSF? Some of the authors have argued the opposite in their recent paper on Xpert Ultra so you can't have it both ways.

If applicable, is the statistical analysis and its interpretation appropriate? Yes
Are all the source data underlying the results available to ensure full reproducibility? Yes may negate the need for tests.
I agree there is seldom a perfect test and clinical prediction rules can be extremely useful in areas where access to tests is limited. I look forward to seeing a CPR that can accurately distinguish TBM from other HIV-related brain infections / pathology and commend researchers in this pursuit.
Discussion-Para5, Do the authors really think there are dead bacilli in CSF? Some of the authors have argued the opposite in their recent paper on Xpert Ultra so you can't have it both ways.