Machine learning's potential for rapid LRTI diagnosis

By Hugo Francisco de SouzaReviewed by Sophia CoveneyJul 17 2023

In a recent preprint study posted to Preprints with The Lancet*, a team of researchers evaluated the use of prediction models along with clinical information, metatranscriptomics, and lower respiratory tract microbiome.

Their results suggest that machine learning models may become a rapid diagnosis tool, circumventing morbidity and mortality associated with conventional microbiological testing.

*Important notice: SSRN publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

LRTI diagnosis

Lower respiratory tract infections (LRTIs) are responsible for over 3 million deaths per year, making them one of the leading infectious causes of human mortality globally. The high morbidity and mortality of LRTIs have historically been attributed to conventional respiratory infection diagnosis. Traditional diagnosis lacks sensitivity, cannot identify 60-70% of causative agents, and takes 24-48 hours or more for infection characterization.

LRTIs are known to exhibit extreme variety and variability in their symptomatic presentation, many of which overlap with non-infectious conditions like asthma, chronic obstructive pulmonary disease (COPD), or cystic fibrosis. Clinicians hence prefer to delay their diagnosis of a patient or risk disease misdiagnosis, both of which could be life-threatening.

Recent studies challenge the classical view of LRTI pathogenesis – traditional knowledge assumes that the lungs are initially sterile. It takes a critical volume of pathogenic microbes invading the lungs to overwhelm the immune response, resulting in rapid infection.

A growing body of research using microbial genomes proposes that LRTIs originate due to a combination of low microbial species diversity, high overall biomass, and host inflammation response.

Alternations in respiratory tract microbiomes have also been observed in non-infectious diseases like asthma, flagging microbiome studies as critical in LRTI identification and characterization. An emerging field, metagenomic next-generation sequencing (mNGS) is being tested as a viable, rapid, and sensitive alternative to traditional diagnosis tools.

mNGS requires microlitre volumes of patient samples and could yield accurate diagnoses in minutes to hours versus the days that conventional diagnostic tools currently take.

About the study

In the present preprint study, researchers attempted to collate and combine respiratory microbiome and host transcriptional profiling with clinical data. They then trained a machine-learning model and tested its diagnostic speed and accuracy when fed with the collated data.

Researchers began by enrolling patients suspected to have LRTIs from the Peking University People’s Hospital, Beijing, between May 2020 and January 2021. After screening for radiography, clinical presentation, and demographic characters in line with the US Centers for Disease Control/National Healthcare Safety Network (CDC/NHSN), 136 participants were chosen for the study.

All participants received traditional microbiological and serological testing for LRTI diagnosis. Researchers additionally collected bronchoalveolar lavage fluid (BALF) for characterization and model training. BALF was sequenced for both DNA and RNA. RNA reads were screened against the human transcriptome and against the SILVA rRNA database to ensure that the remaining reads belonged to the lung microbiome.

Host transcriptome and microbiome were correlated and standardized by comparing transcripts per million (TMP) expression in hosts to the relative concentration of microbial flora. This data was then used to train machine learning models.

Researchers vetted 11 identifying variables from clinical indicators, microbial flora abundance, and host TMP upregulation. Random forest models were utilized, using 91 participants’ data for algorithm training and 45 for testing.

Study findings

Of the 136 patients enrolled in the study, 81 were found to have LRTIs and formed the LRTI cohort, while the rest 55 were placed in the non-LRTI cohort. LRTI-positive individuals were found to have a significantly higher quantity of prior antibiotic use when compared with their non-LRTI counterparts.

Notably, laboratory findings, including white blood cell (WBC) count and inflammation indicators, did not differ between the two groups. This highlights the low characterization power of conventional diagnostic tools.

Patients were LRTIs were found to have significantly reduced BALF microbiota diversity compared to non-LRTI samples. The relative abundance of microbiota was also different between the groups, with BALF of LRTI samples depicting the high abundance of pathogenic Klebsiella pneumoniae, Stenotrophomonas maltophilia, Pseudomonas aeruginosa, and Streptococcus pneumoniae.

In contrast, BALF of the non-LRTI group showed the highest abundance of Halomonas pacifica, a symbiont ordinarily present in healthy lungs and respiratory tracts. The pathogenic microbes in the LRTI samples were either absent or found in trace quantities in the non-LRTI group.

Transcriptome analyses revealed 674 differentially expressed genes (DEGs). Of these, BALF of the LRTI cohort revealed that 613 DEGs were up-regulated, while the remaining 61 were down-regulated compared to the non-LRTI cohort. Screening against the Kyoto Encyclopedia of Genes and Genomes (KEGG) revealed that LRTI up-regulated DEGs were associated with pathogen infection pathways.

Correlations between microbiota diversity and host transcriptomes suggest that 31 host genes (and their relative expression levels) are associated and vary depending on the ratio of normal LRT flora to pathogenic microbes.

Training the Random forest model using this data allowed for predictions of LRTIs using 70 features (11 clinical, 39 lung microbiome, 20 host response). The diagnostic accuracy of the model was found to be 88.2%, and results were obtainable in a few hours, both significant improvements over traditional diagnostic approaches.

The fundamental limitations of this study are that mNGS is currently extremely expensive and requires high technical requirements. Furthermore, while these machine learning models might serve as diagnostic indicators of LRTI, they in no way characterize or explain the pathways or biological functions of the microbiota-host transcriptome interactions observed.

Conclusions

This pre-print study represents a novel approach to lower respiratory tract infection diagnosis. Traditionally, LRTI diagnoses can take multiple days and depict low sensitivity to over 60% of infectious agents. These aspects result in disease misidentifications and intervention delays, significantly contributing to morbidity and mortality.

In this study, researchers characterized microbial abundance in LRTI and non-LRTI cohorts, which they clubbed with host transcriptome and response data. These data were used to train machine learning models, which were subsequently able to correctly diagnose 88.2% of patients with LRTI in a fraction of the time conventional techniques take.

This research, if validated during peer review, and developed to reduce its inherent high cost, could help clinicians rapidly and accurately diagnose LRTI in the future, thereby reducing the high mortality associated with the disease.

*Important notice: SSRN publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.