Artificial Intelligence uses in COPD

In this week’s Newsletter, we explore the role of artificial intelligence (AI) in pulmonary medicine, focusing on COPD screening, diagnostics, and risk prediction.

Our first article explores how AI is reshaping early COPD screening at primary healthcare institutions, making it more accessible and accurate, especially in resource-limited settings.

The second article examines AI’s practical applications in pulmonary diagnostics, from automated interpretation of lung function tests to cutting-edge imaging analysis.

Lastly, our SHAP model explainer demystifies how AI generates its predictions, helping clinicians understand the “why” behind AI-driven insights. Read on to discover how these tools can support you in delivering proactive, patient-centered respiratory care.

AI in Early COPD Screening: Supporting Primary Care Clinicians

Artificial intelligence (AI) is revolutionizing COPD care, offering practical solutions for early detection and personalized treatment planning. In China alone, COPD affects nearly 100 million adults yet remains underdiagnosed due to low awareness and limited access to screening. AI is helping bridge these gaps, making COPD screening more accessible and accurate for primary care providers.

Why Early COPD Screening Matters

COPD is a progressive lung disease that causes persistent respiratory symptoms and airflow obstruction. Detecting COPD early allows for timely interventions that slow disease progression and improve patient outcomes. However, standard COPD screening tools like spirometry require equipment and training, which may only be available in some primary care settings, especially in rural areas. AI enhances screening by offering new ways to identify high-risk individuals and support early diagnosis.

AI-Powered Diagnostic Tools for COPD

AI makes COPD screening easier for primary care providers by interpreting complex lung tests and imaging data. For example, AI software can assist with spirometry interpretation, often matching or exceeding the accuracy of pulmonologists. In one study, AI-assisted spirometry achieved an accuracy rate of 82%, compared to 44.6% for some manual interpretations. This makes it possible for general practitioners to confidently identify patients who may need further testing, even without specialized pulmonary training.

Similarly, AI can analyze lung CT scans to detect early signs of COPD that are difficult to see with the naked eye. For example, a Graph Convolutional Network (GCN) model trained on lung imaging data accurately identified early-stage COPD cases in a study of lung cancer screenings. With these tools, primary care providers can make more accurate referrals and diagnoses, catching COPD earlier and potentially improving patient outcomes.

Using AI to Identify High-Risk Patients from Health Data

Beyond diagnostics, AI also supports COPD screening by analyzing patient data to identify high-risk individuals. For instance, AI systems can review electronic health records to find patients with factors like smoking history or respiratory symptoms and then flag them for further testing. Some AI tools even integrate data from wearable devices that monitor breathing patterns, allowing clinicians to track changes remotely. These devices, which assess metrics such as breathing rate and depth, can detect early signs of respiratory issues, prompting timely medical intervention.

AI-powered screening questionnaires add another layer of accessibility. For example, the COPD Screening Questionnaire (COPD-SQ) used in China asks about symptoms, smoking history, and exposure to pollutants. In a study, this tool demonstrated an accuracy rate of over 70%, making it useful for primary care providers looking to assess risk without specialized equipment

.

Addressing the Challenges of AI in COPD Screening

Despite AI’s benefits, integrating it into primary care poses challenges. Ensuring data privacy is essential, as medical records contain sensitive patient information. In China, new laws like the Personal Information Protection Law set strict standards for medical data use, requiring secure data handling. AI models also depend on high-quality, representative data; models trained on diverse patient data perform better across various demographics, which is crucial for COPD screening in areas with different risk profiles.

Additionally, AI models, especially deep learning ones, can be “black boxes” where the decision-making process isn’t easily explained. Explainable AI tools, like SHAP (SHapley Additive exPlanation) values, offer a solution by showing how specific factors influence a prediction. This transparency helps clinicians understand why certain patients are flagged as high-risk, making AI recommendations easier to interpret and explain to patients.

Building a Stronger AI-Integrated COPD Care System

Primary care providers, data scientists, and policymakers must collaborate to fully realize AI's potential in COPD screening. This includes training clinicians to use AI tools confidently, investing in robust data infrastructure, and promoting data-sharing to improve AI accuracy. Interdisciplinary cooperation can refine AI models and develop screening tools that are accessible, accurate, and equitable.

For clinicians, AI tools are becoming valuable allies in COPD management, providing practical solutions that make early detection possible even in resource-limited settings. By reducing diagnostic barriers and identifying high-risk individuals sooner, AI offers a promising way forward in the battle against COPD, empowering primary care providers to deliver personalized, effective care.

for further reading on this subject see:

AI in Pulmonary Medicine:

Transforming Care for Asthma and COPD

By: Campion Quinn, MD

Artificial intelligence (AI) is making significant strides in pulmonary medicine, offering new ways to improve outcomes for chronic obstructive pulmonary disease (COPD) and asthma patients. Through AI-driven analysis of electronic health records (EHRs), researchers can predict hospital readmissions, customize treatment, and reduce healthcare costs. A recent study by Lopez et al. demonstrated the potential of machine learning (ML) and deep learning (DL) models to identify high-risk patients based on specific clinical features. This approach could redefine how physicians treat chronic respiratory diseases, moving from reactive to proactive care that mitigates risks before they escalate.

AI's Predictive Role for High-Risk Patients

The Lopez et al. study analyzed EHR data from over 5,700 asthma and COPD patients and applied several AI models to predict which patients would likely be readmitted. The standout model, a deep learning multilayer perceptron (MLP), achieved the highest accuracy by balancing sensitivity and specificity, meaning it could correctly identify patients at risk without generating excessive false positives.

What makes the MLP model particularly useful is its ability to analyze a range of EHR data points, such as demographics, lab results, and previous treatments, to create a nuanced picture of risk. For instance, high white blood cell counts and low platelet counts in COPD patients, both of which are indicators of inflammation, were found to be strong predictors of readmission. Similarly, elevated eosinophil counts were associated with higher readmission risk among asthma patients. By capturing these subtle variations, the MLP model enables pulmonologists to anticipate which patients may experience exacerbations, allowing for timely, preventive intervention.

The Use of Medications and Comorbidities in Predictions

Medication history also plays a central role in the AI’s predictive capabilities. The study found that certain medications, such as systemic steroids and inhaled corticosteroids, were among the top predictors of readmission. Asthma patients who required systemic steroids or inhaled corticosteroid/long-acting beta-agonist (ICS/LABA) combinations during their initial hospital stay were more likely to be readmitted. COPD patients using long-acting muscarinic antagonists (LAMA) were similarly at risk. AI can flag patients needing closer follow-up, therapy adjustments, or more aggressive outpatient management by integrating data on the types and combinations of medications administered.

Additionally, comorbidities such as coronary artery disease (CAD) and diabetes added predictive value to the model, particularly for COPD patients. For instance, COPD patients with both diabetes and congestive heart failure showed higher readmission rates. By pinpointing the complex interplay between pulmonary conditions and other chronic diseases, AI can assist pulmonologists in recognizing patients who may benefit from a more comprehensive, multidisciplinary approach to care.

SHAP Values: A Window into AI’s Predictive Process

An exciting aspect of the study was the use of SHapley Additive exPlanation (SHAP) values to break down the MLP model’s predictions, offering clinicians insights into how specific features impact risk assessment. SHAP values reveal which variables most strongly influence the AI’s output, thus enhancing trust and transparency in its predictions. For example, low mean platelet volume was linked to increased readmission risk for COPD patients, while high levels of certain white blood cells had the opposite effect. By visualizing these factors, pulmonologists can better understand the "why" behind AI predictions and make informed decisions to address risk factors directly.

Bridging Disparities with AI-Driven Stratification

Beyond clinical features, the study highlighted AI’s potential to address healthcare disparities. Asthma and COPD patients from Black and Hispanic backgrounds had notably higher readmission rates, a trend likely influenced by socioeconomic and environmental factors. The AI model’s ability to stratify risk across demographic groups can inform healthcare providers where to focus their efforts to close health gaps. However, the study also acknowledged the risk of algorithmic bias, especially when models are trained on data that may not fully represent all demographic groups. To counteract this, fairness-aware AI models are essential to ensure that predictive insights are accurate and equitable.

Integrating AI Predictions into Clinical Practice

While AI provides substantial pulmonary care benefits, its integration into day-to-day practice poses some challenges. EHR systems often lack standardized data entries, and complex models like MLP require continual updates and retraining to remain accurate. Additionally, securing sensitive patient information is paramount to maintaining compliance with healthcare regulations.

Despite these hurdles, AI's benefits for managing chronic respiratory diseases are compelling. With AI's support, pulmonologists can gain near-real-time insights into a patient's condition and allocate resources more efficiently. Imagine a future in which patients identified as high-risk by AI are automatically assigned follow-up appointments within days of discharge, receive home-based support, or are prescribed personalized medication adjustments—all to prevent subsequent hospital admission.

The Future of AI in Pulmonary Medicine: Bridging Gaps, Enhancing Care

The study by Lopez et al. illustrates that AI's role in pulmonary medicine extends far beyond prediction; it offers a roadmap for personalized, data-driven care. AI models such as MLP identify high-risk patients and provide actionable insights that could help physicians optimize care. As EHR integration becomes more seamless, the potential for AI-driven support in clinical decision-making will only grow, ultimately leading to a new standard of care that is both personalized and proactive.

The ongoing collaboration between clinicians, AI researchers, and healthcare policymakers is essential to realizing this potential. By prioritizing ethical AI design, investing in infrastructure, and fostering transparency, we can look forward to an era where AI not only enhances pulmonary care but does so in a way that is accessible and equitable for all patients. AI-driven predictive models may soon become indispensable tools, transforming pulmonology from a reactive specialty to one that anticipates and addresses health risks before they become critical.

SHAP Values: Unlocking AI’s Black Box

By: Campion Quinn, MD

As artificial intelligence (AI) becomes more integrated into healthcare, it offers clinicians powerful tools to support decision-making. However, one of the biggest challenges with AI models is that they can be like “black boxes”—we often don’t know exactly how they make their predictions. This lack of transparency can make it difficult for physicians to trust or rely on AI outputs. SHAP values (SHapley Additive exPlanations) are a tool designed to break down model predictions and show how specific factors influence each outcome.  (FYI: SHAP values are named after the American mathematician Lloyd Shapely, who introduced the concept of cooperative game theory in 1951.) Understanding SHAP values can make AI predictions clearer and more actionable for clinicians, supporting more personalized and targeted patient care.

What Are SHAP Values and How Do They Work?

SHAP values are based on a game theory concept (a mathematics field studying cooperative decision-making). In simple terms, SHAP values measure the impact of each feature (e.g., age, lab results, medications) on a prediction. Think of each feature as a “player” on a team, each making a different contribution to the final score. SHAP values assign each feature a “contribution score,” showing whether it raises or lowers the prediction and by how much.

For instance, imagine an AI model predicting the readmission risk for COPD patients. The model examines features like lab results, medication use, and age to produce a risk score. SHAP values then break down that score, showing how much each feature contributes. For example, a high white blood cell count might increase the risk score (a positive SHAP value), while fewer recent hospital visits could lower the risk (a negative SHAP value). This level of transparency allows clinicians to see precisely what’s driving a prediction and where their attention may be most needed.

Why SHAP Values Matter for Clinicians

In healthcare, SHAP values are valuable because they make AI models less mysterious and more understandable. Physicians often hesitate to act on an AI prediction without knowing its rationale. SHAP values provide this rationale, revealing why a model flags a patient as high-risk and which factors contribute to that assessment. This allows clinicians to feel confident in their decision-making and supports patient-centered care.

For example, suppose an AI model identifies a COPD patient as high risk for readmission. In that case, SHAP values might show that factors such as recent steroid use, inflammatory markers, or diabetes are driving the risk. A clinician could then address these factors by adjusting steroid dosages, monitoring inflammation more closely, or optimizing diabetes management.

Examples of SHAP Values in Action

Consider these practical scenarios for SHAP values in clinical decision-making:

1. Lab Values: SHAP values might reveal that elevated eosinophil counts are a significant factor for a patient at high risk of readmission. This could prompt a clinician to adjust anti-inflammatory treatments or schedule closer follow-up visits.

2. Medication History: If recent steroid use is a significant factor in the risk score, the clinician might reduce the steroid dosage (if appropriate) or implement additional measures to manage side effects, reducing the likelihood of readmission.

3. Comorbidities: For a COPD patient with diabetes, SHAP values might indicate that high blood glucose is a significant contributor to readmission risk. This would guide the clinician in strengthening glucose management as part of the care plan.

How SHAP Values Benefit Clinical Practice

SHAP values do more than clarify AI predictions—they empower clinicians to take targeted actions. If certain factors like lab markers or medications are shown to contribute significantly to a patient’s risk score, a clinician can use this information to make precise intervention decisions. SHAP values also help clinicians validate AI predictions, showing if the highlighted factors align with known risks. If they do, it builds confidence in the AI tool; if not, it may prompt further investigation or adjustments.

Additionally, SHAP values promote fairer healthcare by identifying disparities in AI predictions across demographics. This can help ensure that AI tools serve all patients equitably, a critical factor in healthcare.

Conclusion

For clinicians, SHAP values are vital to understanding AI’s decision-making, making these tools both interpretable and trustworthy. By revealing the “why” behind AI predictions, SHAP values allow physicians to confidently apply AI insights, creating more personalized and equitable care for their patients.