This proof-of-concept study aimed to evaluate the performance of a non-invasive ultrasound-based technique for the screening of meningitis in neonates and infants with a permeable fontanel using DL models to classify patients based on their WBC count in CSF. It has shown a sensitivity of 100% and a specificity of 90% at classifying patients to meningitis and control classes, with only one control misclassified. While the sample size is relatively small, the device takes multiple HR images from each patient as CSF naturally flows beneath the fontanel, resulting in a large set of different images (781) to train the algorithm.
The availability of precise screening tools is necessary for accurately modulating suspicion of ABM and effectively targeting invasive confirmation procedures, such as LPs. Efficient identification of meningitis cases remains a challenge in LMICs, where scarce resources often result in underdiagnosis. This underscores the urgent need to improve screening strategies to identify potential cases of ABM, as delayed diagnosis and treatment significantly increase morbidity, sequelae, and mortality rates24-28. In this scenario, a non-invasive screening tool for ABM could help identifying the cases that need to be referred to a health facility with the capacity to perform LPs, to prioritize the scarce available results, and, if necessary, start empirical treatment when no other options are available. Conversely, HICs require targeted screening tools to identify patients at risk of ABM, particularly in the youngest pediatric population. The highly unspecific clinical presentation in this age group and its high associated mortality necessary lead to a proactive screening approach, resulting in a high number of LP indications, but a low yield of positive results14,19. In fact, it is remarkable that 80% of patients in our cohort underwent a LP due to unspecific presentations such as fever and elevated inflammatory markers in blood along with their young age14,19. Three confirmed cases (and no controls) presented neurological signs, which reinforces the message that meningitis needs to be ruled out in all symptomatic patients. All three also presented macroscopic ultrasound signs of infection, such as meningeal enhancement, and two of them had ABM-associated ventriculitis. Of note, basal ultrasounds were performed within the context of the study and are not necessarily included in the usual evaluation of neonates with suspected meningitis worldwide (as these exams are not always available). These observations are consistent with neurological manifestations presenting in more advanced or severe cases of neonatal meningitis. In contrast, the remaining three cases of meningitis did not exhibit any neurological signs or ultrasound abnormalities, highlighting the need for conservative management protocols with the currently available resources. Nevertheless, the introduction of precise screening tools for ABM could significantly enhance the accuracy of initial suspicion and improve management strategies in such cases. The device aims to provide the same information as laboratory cell count in CSF, but cannot substitute microbiological tests. While all positive cases would still need a LP to be performed to try to identify the etiology of meningitis, many LPs could potentially be avoided by using this non-invasive technique. Among immunocompromised patients for whom a normal cell count in CSF cannot exclude meningitis, LPs should be routinely done for microbiological testing as usual. In the scenario of early stages of symptoms, when pleocytosis can still be absent, the device would provide the same information as cell count, therefore meningitis should not be excluded with a negative result. However, the capacity of measuring WBC in CSF as often as necessary with no risk for the patient, would give clinicians more tools to monitor such cases, besides of repeating LPs as recommended in the literature 29. Furthermore, the device could be used in very ill patients in whom LPs are contraindicated, helping to modulate the suspicion of meningitis before clinical stabilization.
This proof-of-concept study shows that the combination of HR technology and DL provide a reliable screening tool for meningitis. We have achieved a high level of accuracy in measuring CSF WBC levels above and below 30 cells/mm3, which represents valuable information to modulate the initial level meningitis suspicion. While this will be validated in a feasibility study, this proof-of-concept study shows that the technique provided a reliable screening tool for meningitis, as we confirmed 100% of cases, and applying our method in real-life screening of patients, could have potentially spared LPs in up to 90% of controls. A more accurate meningitis screening, which modulates clinical suspicion, can lead to a more accurate indication for LP and an earlier initiation of treatment, which would result in better patient outcomes and in a reduction of the risk of long-term sequelae6-8.
Three LPs were hematic (with >1000 RBC/mm3) needing for WBC correction, similarly to previously reported14, challenging laboratory interpretation of results. By using correction formulas to define the allegedly real number of WBC/mm3 in CSF (or ground truth), the device has correctly classified all of them, reinforcing previous results regarding these formulas to have good accuracy to diagnose children with meningitis and hematic LPs21. This non-invasive tool would give non-blood-contaminated information to clinicians of WBC in CSF, which is particularly interesting for younger infants, those typically more prone to hematic punctures13. Our device may help monitoring meningitis cases and their WBC decrease in CSF after starting appropriate treatment. In fact, in the meningitis patient with two LPs and image acquisitions performed, we already saw a decline in WBC by HR, and that non-invasive exercise could be done as many times as necessary until normalization of CSF. Finally, within our cohort we had 3 controls with pleocytosis who were not considered pathological by their clinicians30. Having access to a non-invasive tool for sequential monitoring and confirmation of the low degree of suspicion for meningitis would offer objective information and greater certainty in these grey-area scenarios.
While all meningitis patients were classified correctly with our DL algorithm, 1 out of 10 analyzed control patients was misclassified as belonging to the meningitis class. This patient had the lowest number of HR image frames to analyze (9 frames, while the overall average number of used HR image frames was 46). Furthermore, that patient had a very limited subarachnoid space (less than 2mm), which made us consider the option of having artifacts from surrounding structures, but when analyzing the results from patients with similar CSF space, the rest of them were all correctly classified. Therefore, we suspect that the low number of images was not representative enough and could lead to the misclassification. Further investigations are needed though to confirm these suspicions, and a larger number of acquisitions/patient has been implemented in our phase II investigations. Another limitation of our study is that 9 initially enrolled patients did not present good quality HR images to be analyzed by our DL techniques due to wrong locations of image acquisitions or blood vessels interfering in the signal (Figure 1, supplementary material). Of note, frames excluded for the classification algorithm training presented in this manuscript, were used to train background algorithms (not presented here). In fact, those images have helped building the background algorithm we are using to locate the good structures in a LR image to avoid vessels or wrong anatomical locations to optimize HR images acquisition. However, the more participants were enrolled the better the images were, which indicates that most low-quality images were discarded due to the usual learning process of both clinicians and engineers for a novel technique. The possibility of taking dozens of images per patient at every acquisition in a short time, allows us to rapidly gain experience and provide our algorithms with a steep learning curve, both to exclude bad quality images automatically, and to classify good quality images as meningitis or controls. For instance, the results achieved with this relatively small sample size, have already helped us develop significant improvements in the device currently being used in the following, automatized, phases of the study.
The fact that 84% of 781 HR images and 94% of individuals have been successfully classified in a relatively small sample of patients is an encouraging indication of the robustness of the tool as a non-invasive technique to help clinicians screen for meningitis. A feasibility study with a higher sample size is already being conducted. As part of future steps, we will provide clinicians with quantitative results with high sensitivity at low concentrations (5 cells/mm3), instead of the current binary classification (meningitis yes/no). A limitation of the study is that we have used a threshold of 30cels/mm3, which has not misclassified any patient in this specific cohort, but a quantitative number will allow clinicians to interpretate the results according to their thresholds, which vary across hospitals and age groups, and in early stages of the disease. Also, we are working on the differentiation between lymphocytes and neutrophils to distinguish bacterial from viral meningitis due to management, therapeutic and prognostic implications. Furthermore, new versions of the device will be completely automatized, easily portable and, therefore, feasible to be used even in remote rural areas in LMICs. Finally, this is a first-in-class device, and further applications of the technology to detect WBC in other serous fluids (such as in cases of uveitis or peritonitis) are being developed.
In conclusion, our non-invasive ultrasound-based technique, utilizing DL models to determine WBC in CSF, shows great promise in screening for meningitis in neonates and infants with a permeable fontanel. Although acquiring CSF samples poses a challenge, the ability to correlate their laboratory WBC counts with numerous HR images per patient non-invasively empowers us to advance quickly in training DL algorithms. This highly sensitive and specific technique has the potential to aid clinicians in accurately identifying potential cases of meningitis, particularly in patients with nonspecific clinical presentations. The use of our device could ultimately lead to a more agile meningitis screening, and modulate LPs indications among the youngest. This will allow a more efficient use of resources and improve patient outcomes at both the individual and public health levels.