This article delves into the use of Artificial Intelligence (AI) in respiratory medicine, focusing on its applications in diagnosing and managing lung conditions such as lung cancer, asthma, COPD, and interstitial lung disease, using advanced techniques like machine learning and deep learning.

---

The term “artificial intelligence” (AI) was coined by Professor John McCarthy in 1955. He defined AI as “the science and engineering of making intelligent machines.”

AI involves simulating human intelligence in machines designed to think and learn in ways similar to humans. This encompasses creating algorithms and software capable of executing tasks that usually necessitate human intelligence, including visual perception, speech recognition, decision-making, and language translation.

---

Branches of artificial intelligence

AI can address real-world challenges through the following methods and techniques:

1. Machine learning (ML)
2. Deep learning (DL)
3. Natural language processing (NLP)
4. Robotics
5. Expert systems
6. Fuzzy logic

Machine learning

Machine learning is a branch of AI that focuses on training algorithms to learn from data and make predictions or decisions autonomously, without being explicitly programmed for specific tasks.

Deep learning

DL is a specific type of ML that is based on artificial neural networks (ANNs) inspired by the structure and function of the human brain and are composed of layers of interconnected “neurons” that process information. The term "deep" signifies that these networks consist of numerous layers.

Natural language processing

NLP focuses on utilising natural human language to extract information for communicating with machines. It employs computational methods to process and analyse human language data—such as text, speech, and writing—so that computers can comprehend and respond to human language in a manner akin to human interaction.

Robotics

Robotics, a subset of AI, encompasses designing, building, operating, and utilising robots. This interdisciplinary field merges AI with hardware and engineering to develop machines, or robots, capable of executing tasks typically associated with human intelligence.

Fuzzy logic

Fuzzy logic provides a mathematical system for addressing uncertainty and imprecision in decision-making. Unlike classical (binary) logic, which operates solely on a strict yes or no basis, fuzzy logic permits decisions to be made according to varying degrees of truth, expressed as values ranging from 0 to 1.

---

Applications of artificial intelligence in respiratory medicine

Lung cancer‑radiology

Numerous research studies have indicated that early identification of lung cancer through low-dose computed tomography (LDCT) can decrease mortality rates by approximately 20%. The US Preventive Services Task Force has determined that annual LDCT screening for lung cancer offers a moderate overall benefit. While LDCT aids in detecting lung cancer early among high-risk groups, it also yields false-positive outcomes, leading to unnecessary examinations and invasive procedures. Moreover, due to human visual limitations and variations in interpretation among radiologists, as many as 35% of lung nodules, particularly small ones, may be overlooked during initial screenings.

In the 1960s, the initial effort to utilise a computer for lung cancer detection was made. A computer-aided diagnosis (CAD) system was created to identify pulmonary nodules in chest X-rays.

AI technology is being increasingly utilised in LDCT scans for lung cancer screening. A team of bioengineers has made use of AI technology to convert low-dose CT images into images of superior quality by post-processing the images. Thus, AI can produce high-quality images from low-dose CT scans, which typically have poor resolution and noise artifacts, eliminating the need for high-dose radiation CT scans.

An AI trained using an adequate quantity of radiology images can localise and identify the lesions in digital images of chest X-rays and CT scans with reasonable accuracy, sometimes better than radiologists. Coupled with histopathology images, AI can accurately diagnose lung cancer, assist in treatment decisions, and predict the prognosis. AI technology not only surpasses radiologists in accurately identifying lesions from CT images but also demonstrates a commendable accuracy of over 80% in distinguishing between benign and malignant lesions.

Lung cancer‑histopathology

Researchers have effectively utilised AI, including machine learning (ML) and deep learning (DL), to not just detect different cancer types but also forecast cancer recurrence based on digital histopathology images. Furthermore, AI can differentiate between malignant and premalignant lesions.

High-quality digital images of pathology slides, commonly stained with H and E stains, are analysed by an AI algorithm. The AI, trained using deep learning on a vast dataset of normal and pathological slide images, detects the region of interest (ROI). A well-trained AI can discern whether abnormalities within the ROI are benign or malignant. Additionally, if malignant, it can also classify the subtype of lung cancer.

Lung cancer‑management

AI can be used to map the surgical site in patients undergoing surgical resection of lung tumours and can also determine whether patients will need adjuvant chemotherapy. For patients undergoing radiotherapy, AI shows promise in forecasting the onset of post-radiation pneumonitis, although its current predictive accuracy remains somewhat constrained. Additionally, in the realm of lung cancer, researchers are employing AI to anticipate treatment outcomes and patient survival rates.

Interstitial lung disease

Among high-risk groups like individuals with familial idiopathic pulmonary fibrosis (IPF) and rheumatoid arthritis, employing an automated computer algorithm to analyse radiological abnormalities in high-resolution CT scans has been proven to objectively enhance the diagnosis of interstitial lung disease (ILD). Numerous researchers have utilised AI for screening, diagnosis, prognostic prediction, mortality assessment, lung cancer development, and therapeutic response in ILDs.

AI-DL has improved the diagnostic precision of chronic hypersensitivity pneumonitis, cryptogenic organising pneumonia, nonspecific interstitial pneumonia, and common interstitial pneumonia patterns. AI has demonstrated good performance in detecting chronic fibrosing ILD in chest X‑rays. A study comparing the performance of an AI-DL algorithm to that of experienced radiologists in classifying ILD using a cohort of 150 high-resolution CT images found that the AI had an accuracy of 73.3% compared to the median accuracy of all radiologists (70.7%), and outperformed 60 (66%) of the 91 radiologists. A meta-analysis of 19 studies that used AI for ILD diagnosis using chest CT showed diagnostic accuracy ranging from 78% to 91%.

Asthma

AI algorithms are being investigated for their potential in asthma screening, diagnosis, phenotype identification, and in assessing and managing asthma control. To diagnose asthma, researchers have combined an AI algorithm with a number of techniques, including forced oscillation techniques to identify airway obstruction, using a wearable sensor to identify wheezing sounds, data from capnography, clinical characteristics, and spirometry data from patient health records.

AI technology was also used in a study where the auscultation recordings of asthma patients were analysed using AI-ML to differentiate asthma patients with and without abnormal breath sounds. AI-DL and ML have also been used to differentiate asthma, COPD, and ACO by analysing the clinical data from patients’ digital health records. AI has also been used to predict the response to corticosteroids in asthmatics.

Chronic obstructive pulmonary disease

AI models have been created to evaluate the severity, prognosis, treatment response, risk of exacerbations, and mortality in COPD patients.

AI can analyse LDCT images from lung cancer screenings to detect COPD-related changes. Additionally, an AI algorithm has been developed to identify COPD patients based on distinguishing low serum levels of N-acetyl-glycoprotein and lipoprotein, achieving an accuracy exceeding 80% when compared to control groups. Using data from the COPD Gene study database, AI has also been used to discover specific COPD phenotypes as well as the genetic and molecular pathways causing disease development in various COPD subtypes.

Prediction of acute exacerbations in chronic obstructive pulmonary disease

Numerous studies have been conducted to determine the severity and predict future exacerbations in COPD patients using AI. This is achieved by feeding the data from patients’ health records, self-reported symptoms by patients, and monitoring certain parameters using sensor-enabled devices, mobile applications, or computer software. Researchers have used AI-ML for the early detection of acute respiratory failure, ventilator dependence, and mortality in patients with COPD after hospitalisation with excellent predictive performance.

Pulmonary function testing

Pulmonary function testing includes spirometry, measurement of lung volumes, and carbon monoxide diffusion capacity of the lungs. Currently, the interpretation of lung function tests is done by health-care workers based on published guidelines. Inter-reader variability contributes to diversity in reporting, potentially due to factors like limited understanding of lung physiology, unfamiliarity with recommendations, training gaps, or oversight. This issue becomes particularly problematic in high-volume environments generating thousands of reports monthly.

The researchers used an AI-DL algorithm to automate the reporting process by analysing 16,502 files in portable document format of spirometry reports with more than 92% accuracy. Shockingly, AI perfectly matched the spirometry pattern interpretations (100%) compared to 74.4% by pulmonologists and gave a correct diagnosis in 82% of cases compared to 44.6% by pulmonologists.

Pneumothorax

The chest X-ray remains the most commonly used radiological investigation in the diagnosis of pneumothorax. However, a small pneumothorax may not be seen on chest X-rays. Up to 20% of them may be missed on chest X-rays but can be detected by CT scans (occult pneumothorax). Researchers have developed various AI models to identify pneumothorax in large sets of chest X-rays, mainly from open-source databases, retrained them using their own customised training dataset, and compared their performance to radiologists’ readings of those X-rays with good results. AI was also used to quantify the pneumothorax by researchers, with promising results.

Pleural effusion

Researchers are investigating AI technology to automatically detect and measure pleural effusion in radiological images, potentially alleviating the workload for healthcare professionals in high-volume environments and enhancing diagnostic accuracy.

Researchers have assessed the effectiveness of AI in identifying tuberculous pleural effusion by training four AI algorithms with a set of 28 features derived from statistical analysis. They compared these algorithms' performances with that of pleural fluid ADA. The top-performing algorithm achieved a sensitivity of 90.6% and a specificity of 92.3% in detecting tuberculous pleural effusion.

SARS‑CoV‑2/COVID‑19 pandemic

Throughout the COVID-19 pandemic, AI was employed by researchers for a range of purposes such as diagnosis, triage, management, contact tracing, and forecasting future outbreaks. In the very early stages of the COVID-19 pandemic, an AI epidemiology algorithm (BlueDot) was used to predict the spread of this novel infection by analysing the data generated by the International Air Transport Association.

Other pneumonias

AI models were successfully trained with X-ray images to automatically detect consolidation. One AI model (InceptionV3) was able to achieve an accuracy of 92.8% with a sensitivity of 93.2%, a specificity of 90.1%, and an AUC of 0.968 for detecting viral pneumonia when compared to normal chest X-rays. The AI model also achieved an accuracy of 90.7% with a sensitivity of 88.6%, a specificity of 90.9%, and an AUC of 0.94 for distinguishing bacteria from viral pneumonia.

Tuberculosis

AI systems were first developed to detect lesions in the chest X-rays of people suspected to have tuberculosis (TB). AI systems detect the preset morphological features of lesions in chest X-rays to screen for TB. An AI system was employed to assess treatment response in TB by analysing chest X-rays taken before, during, and after treatment using a subtraction image analysis algorithm. This method determines if the patient is responding positively to treatment.

AI technology is currently employed in the discovery of novel drugs for TB treatment. Another possibility for the future is that the AI-DL can be used to analyse patients’ clinical, radiological, microbiological, and treatment data to identify risk factors for the development of TB, including any novel gene associated with TB disease, predict adverse drug events, and predict the risk of relapse, drug resistance, and mortality.

---

Disclaimer- The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of M3 India.

About the author of this article: Dr Waseem Ud Din is a practising pulmonologist from Srinagar.