Managing pneumonia with artificial intelligence (AI)

8 April, 2020

As artificial intelligence (AI) becomes increasingly important in the global healthcare system, more scientists are using AI to improve public health efforts in the fight against infectious diseases. The introduction of AI has facilitated clinical physicians in improving the management of pneumonia.¹

In 2017, pneumococcal pneumonia was the leading cause of death due to lower respiratory tract infection (LRI) in children younger than 5 years old, followed by those caused by the respiratory syncytial virus, and Haemophilus influenzae type B.² Meanwhile, in adults older than 70 years old, although death rates due to pneumococcal pneumonia have reduced significantly between 1990 and 2017, death rates due to the respiratory syncytial virus pneumonia have only decreased minimally. The management of pneumonia has also become increasingly challenging with the emergence of multidrug-resistant organisms.¹ As such, the use of AI to support clinical decision-making seems promising.

In clinical practice, the utility of AI is made possible due to the availability of data from electronic health records (HER) as well as advances in computational performance.¹ This signifies the ability of computer systems to operate at high speeds and large capacity to handle information in a relatively short period. Making use of a combination of machine learning (ML) and neural networks (NN), these intelligent systems not only store massive amounts of clinical information but are also capable of predicting various clinical situations.

Current research on AI in pneumonia focuses on diagnosis via the study of chest X-ray (CXR) patterns – largely due to the role of CXR in distinguishing aetiologies of pneumonia.¹ A 93% accuracy rate with 93% sensitivity and 90% specificity was achieved in a study using an NN algorithm for pneumonia detection with a CXR image dataset; the area under the receiver operating characteristic curve (AUROC) was 97%.³ The test was also able to distinguish bacterial pneumonia from viral pneumonia. A different approach using demographic characteristics, signs, symptoms, and comorbidity data to predict pneumonia achieved 47% sensitivity and 97% specificity for pneumonia diagnosis, with AUROC of 87%.⁴ Similarly, other studies have demonstrated that AI algorithm improved the quality and efficiency of thoracic disease diagnosis⁵Studies using AI have also shown positive contributions in choosing empirical therapy of acute pulmonary infections.⁶

In conclusion, AI could enhance pneumonia screening programmes; aid in early radiological diagnoses; and create more efficient referral systems for appropriate management.