TBorNotTB: A novel AI-driven tool to streamline tuberculosis evaluation in hospitals

A new clinical decision support system, "TBorNotTB," helps streamline airborne infection isolation decisions, reducing workload while maintaining TB detection accuracy.

Study: TB or not TB? Development and validation of a clinical decision support system to inform airborne isolation requirements in the evaluation of suspected tuberculosis. Image Credit: SewCreamStudio/Shutterstock.com

In a recent study published in Infection Control and Hospital Epidemiology, researchers presented a clinical decision support system (CDSS) to guide the evaluation of suspected tuberculosis (TB).

Background

In the United States (US), TB incidence has reduced from 10.4 cases per 100,000 people in 1992 to 2.2 cases per 100,000 in 2020, with nosocomial transmission being rare in recent years.

Nevertheless, TB incidence surged to 2.9 per 100,000 people in the US in 2023. Meanwhile, hospitals face capacity and staffing challenges, rendering decision-making for suspected TB patients more urgent.

Current guidelines for suspected TB recommend testing sputum samples for mycobacterial culture and acid-fast bacilli (AFB) smear, with airborne infection isolation (AII) implemented.

Nevertheless, cultures usually take ≥ two weeks, and nearly half of individuals with pulmonary TB show negative smears. While nucleic acid amplification tests (NAATs) have greater sensitivity than AFB smears, they are significantly less sensitive than culture.

As such, discontinuing NAATs and smear results would be premature and risk transmission to healthcare workers and other patients. Therefore, infection prevention and control (IPC) personnel often perform an additional review of patient records before aII discontinuation.

In addition, discussions with relevant consultants and the primary team may follow to ascertain whether TB is unlikely to favor AII discontinuation, which is time- and labor-intensive.

About the study

The study aimed to develop a clinical decision support system (CDSS) to assist in discontinuing airborne infection isolation (AII) while ensuring appropriate evaluation of suspected tuberculosis (TB) cases. Researchers created the "TBorNotTB" system to streamline AII discontinuation without compromising care for patients with possible TB who test negative on nucleic acid amplification tests (NAAT) and acid-fast bacilli (AFB) smears.

A panel of experts designed and refined a set of questions based on clinical guidelines and epidemiologic data using the Delphi consensus method. These questions were tested on hospitalized patients at Massachusetts General Hospital (MGH).

The CDSS assigned scores based on epidemiologic risk factors, TB symptoms, medical history, bronchoscopy/sputum results, and chest imaging. If a patient’s score exceeded a set threshold, AII was automatically discontinued; otherwise, further evaluation was recommended.

To validate the CDSS, researchers conducted a case-control analysis using data from the Mass General Brigham (MGB) system. The tool was retrospectively applied to patients with culture-confirmed TB (cases) and matched culture-negative controls.

Variables that did not improve predictive accuracy were removed or down-weighted. The model’s performance was then fine-tuned by adjusting the weight of key TB predictors and analyzing its sensitivity, specificity, and area under the curve (AUC).

A sub-group analysis compared the CDSS’s effectiveness before and after the COVID-19 pandemic. Additionally, the study estimated the total infection prevention and control (IPC) person-hours saved at MGH annually through the CDSS implementation.

Finally, the study suggests that TB is unlikely to be a strong driver for AII discontinuation, given the time and labor-intensive nature of the process.

Findings

The researchers identified multiple predictors of TB in the case-control sub-study. Cavitary lesions or other findings suspicious of TB on the chest radiology report were strongly associated with active TB.

Further, prior residence in a country highly endemic for TB was the strongest epidemiologic risk factor for TB. A history of positive interferon-γ release assay (IGRA) was also a strong predictor of active TB.

Notably, a negative IGRA or tubulin skin test within the past month showed a negative association with TB. A history of conditions deemed to elevate TB risk was not associated with TB.

A history of weight loss was the only traditional TB symptom that predicted active TB. Further, marked improvement or resolution of symptoms with treatment for alternative diagnosis showed a strong negative association with TB.

This initial CDSS model showed 16% specificity and 100% sensitivity. Following iterative revisions to the scoring system, the final model showed a sensitivity of 100%, specificity of 27%, and AUC of 0.87.

There were no differences in model performance pre- and post-COVID-19. The researchers estimated that TBorNotTB could save more than 40 IPC person-hours annually at MGH.

Conclusions

Together, the study developed and validated a novel CDSS to guide the diagnostic evaluation of patients with suspected TB in low-prevalence settings. Prior positive IGRA, residence in a TB-endemic country, and chest radiological findings suspicious of TB were significant predictors of TB.

The final CDSS model showed modest specificity and high sensitivity in detecting TB. Overall, TBorNotTB could help reduce the risk of nosocomial transmission and save considerable IPC person-time.