Objectives: Surveillance, audit and feedback of invasive mold diseases (IMD) is fundamental to antifungal stewardship (AFS) but current methods are inefficient and restrictive. Continuous prospective surveillance of IMD in all haematology patients has not been possible because the majority of these infections present as a culture negative pneumonia. Artificial intelligence (AI) targeting medical imaging provides an opportunity to identify fungal pneumonia in real-time in order to present actionable data at the point of care. Here we report preliminary results of real-world prospective validation of deep learning based natural language processing (NLP) of chest imaging reports for real-time detection of IMD in haematology patients that is part of a multicentre clinical trial (Clin Trials.gov NCT03793231).
Objective:To validate NLP of chest computed tomography reports for prospective surveillance of IMD in all hospitalised haematology patients. As a validation study, NLP alerts are not being fedback to clinicians but rather the machine is being validated against human surveillance. Presented here are preliminary results from one activated site, Alfred Health, a major quaternary Australian transplant centre.
Methods: Prospective active manual and electronic surveillance of invasive fungal diseases (IFD) among hospitalised patients under the haematology service commenced on 1 December 2018. Hospitalised patients with IFD were identified by research coordinators through active case finding every week. All suspected IFD cases were reviewed again at 30 days and data collected to determine if they met internationally accepted criteria for IFD diagnosis (EORTC/MSG 2008 definitions). Chest CT reports were prospectively processed by NLP hosted on the hospital server and predictions were compared to IFD cases detected by active surveillance.
Results: From 1 December 2018 to 31 April 2018, there were 8 patients with suspected IFD identified. Of these, one case was excluded because it did not meet internationally accepted criteria for IFD, resulting in 7 patients with confirmed IFD associated with the following characteristics: all had IMD; all had pulmonary site of infection; 2 cases were probable/proven; 2 possible cases had a positive Aspergillus PCR on bronchoalveolar lavage. There were 6 cases of breakthrough IMD on antifungal prophylaxis involving posaconazole in 4, liposomal amphotericin in 1 and fluconazole in 1 case. Of 90 chest CT reports screened by fungalAi-NLP over the study period, 6 of the 7 IFD patients were correctly flagged. The case missed by NLP was an intubated patient with probable pulmonary invasive Aspergillosis associated with isolation of A. terreus. The NLP output in this case had probabilities of 37.8% and 41.0% on serial reports which was below the 50% threshold for triggering a positive alert. The radiologist’s conclusion was not definitive in either report. There were 10 (11%) false positive reports. NLP is tuned for fungal pneumonia so may miss extra-pulmonary sites. False positives could be minimised with multimodal analysis combining other sources of data.
Conclusion: Real-time, embedded, NLP of chest imaging reports facilitates surveillance of diagnostically challenging IMD for an entire haematology population. NLP utilises diagnostic imaging readily available in hospitals rather than the electronic medical record. Generalisability of NLP will be enhanced by our multicentre clinical trial.
Full conference title:
- TIMM (2019)