Search

Objectives Tuberculous meningitis(TBM) is a common cause of meningitis in sub-Saharan Africa. CSF PCR with GeneXpert RIF/MTB Ultra is only 70% sensitive for detection of definite/probable TBM. Many infections can mimic TBM. Metagenomic next generation sequencing(mNGS) can detect the whole diversity of infectious microbes, but can be insensitive to TB in CSF. We assessed whether leveraging CSF mNGS to identify infections combined with a machine learning classifier(MLC), based on host transcriptomic data generated by mNGS, could enhance diagnostic accuracy for TBM. Methods Prospectively enrolled 347 HIV-infected Ugandan adults with subacute meningitis: RNA/DNA libraries were made from CSF and deep sequenced. Non-human sequences were interrogated to identify pathogens. A host transcriptomic MLC was developed from human RNA transcripts using 70 cases. The MLC and mNGS reporting thresholds were then tested on 108 blinded cases within the cohort. Results mNGS was 75% concordant(27/36) for detecting TB in definite TBM cases and 59% concordant(30/51) in definite/probable TBM combined. 3 TB and 3 non-TB pathogens were detected in the probable TBM group. In the possible TBM/indeterminant groups, mNGS identified 3 cases of TBM and 17 other pathogens. The combined mNGS and host-MLC displayed 83.3%(5/6) sensitivity, 86.8%(59/68) specificity, with an area under the ROC curve of 0.83(p=0.009). Conclusion mNGS identified an array of infectious TBM mimics, including many treatable and vaccine preventable pathogens. mNGS was 75% concordant with definite TBM. We further enhanced the sensitivity of the CSF mNGS assay by developing the first CSF-based host MLC to discriminate between TBM and its mimics