Predicting Adverse Outcomes Using Machine Learning of COPD Patients in Hong Kong

Predicting Adverse Outcomes Using Machine Learning of COPD Patients in Hong Kong

This study aims to develop predictive models for patients with a diagnosis of COPD at discharge of an index admission on these outcomes using machine learning:

Primary outcome: Early admission

Secondary outcomes:

1. Frequent readmission
2. Composite outcome (Early + Frequent readmissions)
3. Mortality
4. Longstayers

Chronic obstructive pulmonary disease (COPD) is a common, preventable, and treatable disease that is characterised by persistent respiratory symptoms and airflow limitation that is due to airway and/or alveolar abnormalities usually caused by significant exposure to noxious particles or gases and influenced by host factors including abnormal lung development. It was estimated 3.2 million people died from COPD worldwide in 2015 and there was an increase of 11.6% compared with 1990. COPD is the third leading cause of death globally in 2019.

In Hong Kong (HK), the prevalence rates of COPD in the elderly population aged ≥60years were 25.9% and 12.4% based on the spirometric definition of forced expiratory volume in 1 s (FEV1)/forced vital capacity (FVC) ratio <70% and the lower limit of normal of the FEV1/FVC respectively.4 From our recent study on COPD hospital admissions, there are a total of 67,628 COPD admissions Jan 2017 Week 1 to Jan 2020 Week 3 (before the COVID pandemic) and 11,065 admissions from Jan 2020 Week 4 to Dec 2020 Week 4 (during the COVID pandemic). 5 The burden of COPD hospitalizations is significant and it is important to understand the driver of these admissions for developing suitable strategies to solve the problem and improve the health outcomes of patients suffering from COPD.

Early readmission and frequent admissions resulting from COPD are commonly studied hospital outcomes because of the high financial burden to both individual and state and the high usage of public healthcare resources. With the advent of Artificial Intelligence (AI) and Machine Learning (ML), there has been considerable interest on its application to medicine. Recent metaanalysis showed compatibility of these models in predicting COPD outcomes.7 However, few studies have managed to show that AI/ML are superior to traditional statistical modeling methods, AI/ML are interpretable and can be clinically correlated, and AI/ML can have direct clinical application.

This study aims to develop predictive models for patients with a diagnosis of COPD at discharge of an index admission on these outcomes:

Primary outcome: Early admission

Secondary outcomes:

1. Frequent readmission
2. Composite outcome (Early + Frequent readmissions)
3. Mortality
4. Longstayers

The viability and purported superiority of Machine Learning (ML) models as alternatives to traditional statistical learning methods will be assessed. Apart from that top predictors of each outcome of interest would be identified for suggestions of possible interventions that will improve outcomes (i.e. reduce early admission, frequent admission and mortality rates). Clinical scores for deployment in clinical setting will also be developed.

Inclusion Criteria:

• ≥40 years
• Patients are discharged from 2016 -2022
• Discharge Diagnosis: Using the Discharge Diagnosis ICD Codes found in the Primary Diagnosis to determine if a patient has COPD
• Validated against Spirometry results (for patient with a spirometry reading):

Spirometry reading taken from anytime point before. Patient should have Post FEV1/FVC ratio of < 0.7 in any one of the spirometry readings. If Post FEV1/FVC is not available, we will check if patients have a Pre FEV1/FVC value, and will also include patients with Pre FEV1/FVC ratio of < 0.7 in any one of the spirometry readings.

Exclusion Criteria:

• Admission diagnosis due to causes other than COPD