AI advances efficiency in the lab

Dec. 19, 2019

The clinical microbiology laboratory is exploding with new technology. Among the most exciting are the use of artificial intelligence and interpretive algorithms (AI/IA) to assist in the work up of bacterial cultures. Not having sufficiently trained personnel to staff our laboratories around the country, digital imaging along with AI/IA will help fill those gaps. AI/IA will assist the laboratory technologists and technicians with their workload by ‘handling’ negative and insignificant cultures, allowing the technical staff to spend their time and resources on those cultures that need their educated assessments. This article will review and discuss the AI/IA software currently available that make this possible, as well as what will be coming soon.

Chromogenic agars and AI

With the advent of liquid-based microbiology collection devices, automated culture processing instrumentation together with smart incubators and digital imaging, we can advance into the field of AI in the microbiology laboratory with culture interpretations. First, let’s review current AI studies that look at the use of AI with chromogenic agars.

A study by Faron et. al. looked at methicillin-resistant Staphylococcus aureus (MRSA) screening cultures at four clinical sites which comprised over 57,000 specimens and using three different manufactures’ agars. They showed that the sensitivity of the AI software reading of the culture images was 100 percent, as compared to manual image reading by the microbiologists1. This means that the AI software never called a culture negative that manual reading called positive. The AI software also detected an additional 153 positive cultures that manual reading missed.

A similar study looking at vancomycin-resistant Enterococcus (VRE) screening cultures reviewed over 104,000 specimens at three cites using two different manufacturer’s agars2. Again, the AI software showed a sensitivity of 100 percent, as compared to manual image reading, and detected an additional 499 positive VRE cultures that were missed by manual reading.

In addition to screening cultures, AI software has also been used very successfully with chromogenic agars for the detection of group A streptococci (GAS) from throat specimens, as well as the detection of group B streptococci (GBS) from vaginal/rectal pregnancy screening cultures. The study by Van et. al. showed that AI software had a 100 percent sensitivity as compared to manual image reading using GAS chromogenic agar, and that the AI detected additional positive specimens that were missed by manual reading3.

These investigators also compared the detection of GAS using chromogenic agar and AI to the detection of GAS by a molecular assay. Using a composite true positive definition (culture positive with GAS confirmed by MALDI identification and/or PCR x2 positive), the molecular assay had a sensitivity of 96.9 percent, chromogenic agar plus AI software had a sensitivity of 90.6 percent, while manual image reading has a sensitivity of 87.5 percent. Thus, using chromogenic agar with AI software is not only more accurate than manual image reading, it approaches the sensitivity of PCR testing.

Another study looking at GBS culture screening during pregnancy showed chromogenic agar used with AI software had a sensitivity of 95.5 percent as compared to manual image reading, which showed a sensitivity of 90.3 percent and molecular detection sensitivity of 96.8 percent4. The sensitivity of the chromogenic agar plus AI software was comparable to that of detection by molecular techniques.

Two studies using chromogenic agar with AI software have also been presented for the evaluation of urine cultures. The first showed 99 percent accuracy of the AI software in segregating urines into groups with no growth (29 percent of all urine cultures), those with insignificant growth (27 percent), those that contained significant growth of Escherichia coli (3 percent), and those that contained significant growth of another urinary pathogen (2 percent)5. These four categories entailed 62 percent of their urine specimens, 59 percent of these specimens could be reported in batch mode of 30 cultures with one computer click by the microbiologist, and 56 percent required no hands-on time by the staff.

The second study showed that AI software with another urine chromogenic agar had a 99.8 percent sensitivity as compared to manual image reading6. In addition, when AI software plus chromogenic agar was compared to conventional agar with manual image reading, a significant (p<0.01) reduction of 06:23 for positive urine specimen results and 04:48 for negative urine specimen results was observed.

AI and routine urine cultures

But AI is not just for use with chromogenic media. A recent study by Faron et. al. utilized AI software to segregate significant growth in urine specimens plated to standard media7. Briefly, nearly 13,000 urine specimens submitted for bacterial culture from three different sites were plated on sheep blood and MacConkey agars. All specimens were processed using a 1µL loop and images were captured after zero and 18 hours of incubation. The AI software quantitated each plate and reported the specimen as “non-negative” if either plate contained more than 10 colonies (>104 CFU/mL).

Results were then compared to manual interpretation as either positive or negative for pathogens, based on each laboratory’s urine culture policy. All manual (M) positive (P), automation (A) negative (N) cultures were reviewed by a second technologist. Overall, the AI software was highly sensitive with an average sensitivity of 99.8 percent (range 99.7-99.9 percent). These data included 5,678 specimens that were positive by both methods (MP/AP), and only nine specimens that were MP/AN.

Specificity showed an overall rate of 72 percent, which included 5,598 MN/AN specimens and 2,180 MN/AP specimens. The 9 MP/AN discrepancy results were found to fall into two categories. The most common cause for discrepancy (eight of nine cultures) was due to the presence of microcolonies that were counted as positive by the technologist but were programed to be ignored by the software. Allowing the software to take microcolonies into account, all eight of these cultures would have been detected and placed into the AP category.

The one remaining MP/AN specimen was due to a difference in bacterial count near the reporting threshold for this laboratory (threshold of 50 colonies or greater). For this specimen, the laboratory report had a bacterial count of 55 and the AI software counted just under the threshold of 50 CFU (49 colonies). Interestingly, allowing the software to count the microcolonies and using a threshold of 10 CFU/mL for each laboratory resulted in 100 percent sensitivity of the software. The authors found significant utility in the ability to remove negative specimens from the microbiologists’ review queue. In this study, 43.3 percent of all specimens were resulted as MN/AN, so for a laboratory that processes 350 urine specimens each day, the AI software would reduce the work load by 151 cultures/day, which equates to 55,000 urine cultures annually that would not need individual technologist review.

AI and AST

AI software is also under investigation for the ability to read and interpret disk diffusion results with as few as six hours of incubation. The study by Hombach et. al. showed that rapid disk diffusion antimicrobial susceptibility testing (AST) read at six hours, as compared to standard disk diffusion incubated for 18 hours, showed agreement of 97.2 percent, 97.4 percent and 95.3 percent for Enterococcus faecalis, E. faecium and Acinetobacter baumannii, respectively8.

With Pseudomonas aeruginosa the average readability of inhibition zones was 68.9 percent at eight hours with an overall categorical agreement of 94.8 percent.

A second study by Hombach and colleagues showed that the vast majority of zone diameters for Escherichia coli and Klebsiella pneumoniae were readable after six hours of incubation, and reliable reading for Staphylococcus aureus was possible after eight hours of incubation9. These studies demonstrated that early disk diffusion reading is possible, and that the precision of disk diffusion AST results are not hampered by early reading.

Summary

There are also many efficiencies to be gained by using laboratory automation and AI software in microbiology. Laboratories will see a decrease in their cost per test, an increase in their productivity and be able to handle additional specimen workload without the need to increase staffing. AI is revolutionizing the microbiology laboratory by segregating chromogenic agar screening cultures into positive and negative groupings, counting colonies for our quantitative urine cultures, discriminating morphologies on routine bacteriology media and will soon read our disk diffusion AST results in a rapid fashion.

The promise of AI is that this software will move to the next step of automated release of negative cultures, whether that is on chromogenic medium or traditional culture medium. The utilization of AI in microbiology will allow a future where clinical microbiologists can spend their time on more complex cultures that require their expert attention and, at the same time, save valuable resources for the laboratory. 

REFERENCES

  1. Faron ML, Buchan BW, Vismara C, Lacchini C, Bielli A, Gesu G, Liebregts T, van Bree A, Jansz A, Soucy G, Korver J, Ledeboer NA. 2016. Automated Scoring of Chromogenic Media for Detection of Methicillin-Resistant Staphylococcus aureus by Use of WASPLab Image Analysis Software. J Clin Microbiol 54:620-4.
  2. Faron ML, Buchan BW, Coon C, Liebregts T, van Bree A, Jansz AR, Soucy G, Korver J, Ledeboer NA. 2016. Automatic Digital Analysis of Chromogenic Media for Vancomycin-Resistant-Enterococcus Screens Using Copan WASPLab. J Clin Microbiol 54:2464-9.
  3. Van TT, Kenneth Mata K, Dien-Bard J. 2019. Automated Detection of Streptococcus pyogenes pharyngitis using Colorex Strep A CHROMagar and WASPLab Artificial Intelligence Chromogenic Detection Module Software. J. Clin. Microbiol. doi:10.1128/JCM.00811-19.
  4. Timm K, Baker J, Culbreath K. 2019 Clinical performance of the WASPLab AI/IA-PhenoMATRIXTM software in detection of GBS from LIM-enriched cultures plated to CHROMID Strepto B Chromogenic Media. ASM Microbe 2019.
  5. Poutanen SM, Bourke J, Lo P, Pike K, Wong K, Mazzulli T. 2019. Use of Copan’s WASPLab PhenoMATRIX Artificial Intelligence to Improve the Efficiency of Urine Culture Interpretation. ASM Microbe 2019.
  6. Faron ML, Buchan BW, Samra H, Ledeboer NA. 2019. Evaluation of the WASPLab software to Automatically Read CHROMID CPS Elite Agar for Reporting of Urine Cultures. Submitted to J Clin Microbiol.
  7. Faron ML, Buchan BW, Relich R, Clark J, Ledeboer NA. Evaluation of the WASPLab Segregation Software to Automatically Analyze Urine Cultures using Routine Blood and MacConkey Agars. Submitted to J Clin Microbiol.
  8. Hombach M, Jetter M, Blöchliger N, Kolesnik-Goldmann N, Keller PM, Böttger EC. Rapid disc diffusion antibiotic susceptibility testing for Pseudomonas aeruginosa, Acinetobacter baumannii and Enterococcus spp., Journal of Antimicrobial Chemotherapy, Volume 73, Issue 2, February 2018, Pages 385–391.
  9. Hombach M, Jetter M, Blöchliger N, Kolesnik-Goldmann N, Böttger EC. Fully automated disc diffusion for rapid antibiotic susceptibility test results: a proof-of-principle study, Journal of Antimicrobial Chemotherapy, Volume 72, Issue 6, June 2017, Pages 1659–1668.