ID diagnostics: November 2018

A computer algorithm detects the presence of active pulmonary tuberculosis on chest radiograph with better diagnostic accuracy than a panel of thoracic radiologists.  The authors trained a deep-learning-based algorithm to detect active pulmonary tuberculosis on chest x-rays using a set of 54,000 films from people without active TB disease and 6,800 films with active TB, a portion of the latter of which had been labeled and annotated by a panel of radiologists. If I understand the methods section correctly, the final diagnostic algorithm is actually three independent diagnostic “entities,” each of which submits predictions that are then averaged, and what it churns out is a “probability heat map” overlaying the plain film with per-pixel probabilities that each segment of the chest x-ray represents a TB lesion. To validate the algorithm, the authors had it interpret films from six external, multicenter and multinational data sets. The diagnostic performance of the algorithm was compared to that of 5-person groups of non-radiologist physicians, board-certified radiologists, and thoracic radiologists. Finally, the authors assessed the performance of the physicians when using the algorithm as a diagnostic aid.

The algorithm was highly accurate in identifying TB lesions across all six validation datasets. The authors set two probability thresholds to have the algorithm classify a lesion as positive, based on the results of their internal validation with the training data set – one threshold intended to achieve 98% sensitivity, and another intended to achieve 98% specificity. Using the high-sensitivity cutoff, the algorithm’s demonstrated sensitivity in the external validation data sets ranged from 94-100%, while specificity ranged from 91-100%. Using the high specificity cutoff, the sensitivities ranged from 84-99% and specificities ranged from 99-100%. For comparison, the sensitivities and specificities of the physicians’ reads were 72% and 67% (non-radiologists), 90% and 95% (radiologists), and 95% and 93% (thoracic radiologists).  

The algorithms improved the diagnostic accuracy of the physicians: with algorithmic support, the sensitivities and specificities of physicians’ chest film reads for TB lesions improved to 85% and 80% (non-radiologists), 93% and 96% (radiologists), and 96% and 94% (thoracic radiologists). Of course, another way to say this is that the algorithm did better on its own than when it had a fallible meatbag contradicting its reads. I’m pretty sure this is how Skynet starts, but that’s fine. 30418527 

Rapid direct antimicrobial susceptibility testing performed on urine is highly accurate and reduces turnaround time to eight hours.  Traditionally, urine culture starts with identification of the organism(s) in a urine sample, followed by subculture to media with antimicrobial susceptibility testing.  Why not go straight to the antimicrobial susceptibility testing?  This approach, termed direct susceptibility testing, has previously been shown to have >95% agreements with standard methods when the bacteriuria is monomicrobial and high-grade (references: 21696255 7494020 and 26899829). What this paper brings to the table is a direct susceptibility testing method with a drastically reduced turnaround time, achieved by incubating organisms on Rapid Mueller-Hinton media and using an automated plate-reading system.

The authors included urine samples submitted for culture at a single Parisian hospital over a one-year period that demonstrated both pyuria and monomicrobial bacteriuria by microscopy (n=321, mostly E.coli and K.pneumoniae). They performed antimicrobial susceptibility testing by both the rapid direct susceptibility testing (DST) and traditional methods. When the methods gave different results, the DST method was assumed to be in error; errors were classified as “very major” (DST reported an antibiotic as sensitive, traditional method reported as resistant), “major” (DST reported an antibiotic as resistant, traditional method reported as sensitive), or “minor” (one method reported an antibiotic as intermediate and the other as sensitive or resistant).

Of the 321 urine samples tested, 245 (76%) were included in the analysis; the remainder were excluded due to culture results that did not match microscopy (n=32), proved to be polymicrobial (n=28), or were unreadable after 8h incubation due to insufficient growth (n=16). A total 5285 bug-drug combinations were tested, and of these, 98% of results were concordant between the rapid direct and traditional susceptibility testing methods. Of the discrepancies, the majority were minor errors, with <0.5% each of major and very major errors. Nine out the 14 very major errors involved cephalexin, and eight of the 17 major errors involved amoxicillin-clavulanate; the minor errors mostly involved fluoroquinolones, trimethoprim-sulfa, and piperacillin-tazobactam. The authors estimate the additional cost of performing this rapid testing to be $6 per 16-antibiotic panel.

It’s hard to imagine rapid direct susceptibility testing being cost-effective as a routine practice for urine cultures. However, in the subset of patients who are hospitalized with probable UTI, and particularly for those who have some risk factor for antibiotic resistance necessitating use of an empiric broad-spectrum agent (e.g. carbapenem), rapid susceptibility testing could be an important tool for antibiotic stewardship. I’m left wondering, though – why not do PCR testing for antibiotic resistance genes directly on the urine instead? 30386946

Blood cultures obtained while or immediately after a patient has shaking chills are more likely to be positive. This was a prospective study of patients admitted to the inpatient infectious diseases service at a single hospital in Okinawa over a 2-year period. The authors examined the results of initial blood cultures in those patients who were admitted with shaking chills and were later suspected or found to have an infection with a readily culturable bacteria (total n=214). They compared the outcome of the blood cultures among patients cultured within 2 hours of their first episode of shaking chills versus those cultured later, and found that cultures obtained within 2 hours were more likely to be positive (54% vs 38%; p=0.019). In a multivariate analysis accounting for age, site of infection, vitals, and recent antibiotic exposure, early culture was associated with a nearly doubled odds of culture positivity (aOR 1.88; p=0.046); the only more strongly predictive factor was having had multiple episodes of rigors (aOR 3.39; p=0.012).

I want to point out that blood culture timing in relation to the patients’ fevers did not have any impact on culture positivity in this study, a finding that had been previously shown elsewhere. Rigors are the more specific indicator of bacteremia – and particularly gram-negative bacteremia (reference: 9717936). Anyway, the takeaway from this study is straightforward: if your patient is having shaking chills, get a set of blood cultures without delay. 30059771

In patients with VAP, running the BioFire blood culture ID panel on respiratory samples may allow rapid targeting of empiric antimicrobial therapy. In this study, respiratory samples collected for another multicenter European clinical trial were run through the BioFire FilmArray system (a PCR panel for identifying microorganisms directly from positive blood cultures), and those results were compared to the results of the standard microbiologic workup on the specimens. A total 165 patients from 32 sites were included in this study; the most common identified organisms were P.aeruginosa, A.baumannii, and K.pneumonia, collectively representing 70 of 128 isolates. BioFire results were available within one hour, and compared to standard techniques the BioFire had a global sensitivity of 79% and specificity of 98%, with a negative predictive value of 97%. The authors conclude that use of the BioFire PCR panel directly on respiratory samples could allow for rapid optimization of empiric treatment. I have to imagine that this test could pay for itself in the ICU just by cutting down on carbapenem, pip-taz and cefepime use, to say nothing of the long-term positive effects that would have on the ICU’s antibiogram. 29906599

Culturing periprosthetic tissue in blood culture bottles may improve culture yields.  Periprosthetic tissue cultures are traditionally sent in cases of suspected prosthetic joint infection, but have limited sensitivity. In this prospective study out of the Netherlands, investigators enrolled patients undergoing revision arthroplasties, sending tissue samples for both standard bacterial culture (agar and broth) and culture in blood culture bottles (with subculture of results to blood agar, chocolate agar, etc as standard for blood cultures). A total of 113 revision arthroplasties occurred among 90 patients, half of whom met IDSA criteria for PJI and three quarters of whom had acute infections. While the increase in sensitivity from 84% to 93% was nonsignificant, more pathogens were identified in the blood culture bottles than in standard cultures (89 versus 61 pathogens recovered in total 57 infections, p < 0.001). Microorganisms were cultured only from the blood culture bottles in 30% of cases, half of which represented virulent organisms (i.e. likely causative pathogens). Interestingly, the additionally identified virulent pathogens were S. aureus, S. lugdunensis, E. coli, and P. aeruginosa – none of which are particularly difficult to culture. The additionally cultured “low virulence” organisms included other coagulase-negative staphylococci in 11 cases, E. faecalis in two cases, and a Corynebacterium - some of which I’d consider potentially true causes of PJI. In all, this seems like a simple and cost-effective way to improve the rate at which we’re able to provide targeted versus empiric therapy for prosthetic joint infections. 30430376