Unsupervised Machine Learning of Computed Tomography Angiography Features Uncovers Unique Subphenotypes of Aortic Stenosis With Differential Risks of Conduction Disturbances Following Transcatheter Aortic Valve Replacement

This study demonstrates that unsupervised machine learning applied to pre-procedural computed tomography angiography features can identify distinct aortic stenosis subphenotypes, particularly in male patients, which significantly improve the prediction of conduction disturbances following transcatheter aortic valve replacement beyond conventional risk factors.

The Gist

Imagine you are a master carpenter about to install a new, high-tech door (a heart valve) into an old, tricky house (a patient's heart). You know the house has a narrow, creaky hallway (aortic stenosis) that needs fixing. You have a detailed 3D blueprint of the house (a CT scan) before you start.

For years, doctors have looked at this blueprint and measured specific things: "Is the hallway wide enough?" "Is the wood rotting (calcified)?" "How deep should I hammer the new door?" These are like checking individual dimensions. But sometimes, even with all these measurements, the door installation goes wrong, causing a short circuit in the house's electrical system (conduction disturbances), which might require installing a permanent backup generator (a pacemaker).

The Big Idea of This Paper
This study asked a new question: Instead of looking at one measurement at a time, what if we looked at the entire blueprint all at once to see if the house naturally falls into different "neighborhoods" or "types"?

The researchers used a special computer program called Unsupervised Machine Learning. Think of this like a super-smart librarian who looks at 660 different houses and says, "Hey, these 300 houses look like they belong in the 'Short and Wide' neighborhood, while those 300 look like they belong in the 'Tall and Narrow' neighborhood," without being told what to look for.

What They Found
They discovered that men and women have different "architectural blueprints" that matter for this surgery.

1. For Men (The Three Neighborhoods):
   The computer found three distinct types of male heart anatomy:

   • Type M1 (The "Small & Calm" House): These patients had small amounts of "rust" (calcification) on their valves and smaller heart chambers. They were the safest group.
   • Type M2 (The "Short & Wide" House): These patients had a lot of rust, a very wide hallway, but a short vertical space. This is the dangerous group. They were twice as likely to have an electrical short circuit after the surgery compared to the "Small & Calm" group.
   • Type M3 (The "Tall & Wide" House): These patients also had a lot of rust and a wide hallway, but their vertical space was tall. Interestingly, despite being wide, they didn't have the same high risk as the "Short & Wide" group.

   The Analogy: Imagine trying to fit a big, heavy door into a wide doorway. If the doorway is also very short from top to bottom, the door might get stuck or press too hard against the floor (the electrical system). If the doorway is wide but tall, the door has room to settle without causing a short circuit.

2. For Women (The Two Neighborhoods):
   The computer found two types of female heart anatomy. One had more rust and bigger dimensions than the other. However, unlike the men, neither type showed a significantly higher risk of electrical problems. This might be because fewer women in the study had these electrical issues to begin with, making it harder to spot the difference.

Why This Matters
Before this study, doctors were like carpenters measuring just the width of the door frame. They knew that "rust" and "depth" were bad, but they didn't realize that the combination of a wide frame and a short height was the specific "danger zone" for men.

By using this new "neighborhood" approach:

• Better Prediction: Doctors can now look at a male patient's CT scan and say, "Ah, you look like the 'Short & Wide' neighborhood. We need to be extra careful."
• Personalized Plans: If a patient is in the high-risk "Short & Wide" group, the doctor might choose a different type of valve or place it slightly differently to avoid tripping the electrical wires.
• Beyond the Basics: This method adds a new layer of safety on top of the standard checks doctors already do.

The Bottom Line
This paper shows that our hearts aren't just a collection of separate measurements; they are unique architectural structures. By using AI to group patients into these "architectural neighborhoods," we can predict who is at risk of electrical trouble after heart valve surgery much better than before. It's like moving from checking a single ruler measurement to understanding the entire shape of the house before you start building.

Technical Summary

1. Problem Statement

Transcatheter Aortic Valve Replacement (TAVR) is a standard treatment for severe aortic stenosis (AS), yet conduction disturbances (CDs)—including new-onset left bundle branch block (LBBB) and high-degree atrioventricular block requiring permanent pacemaker implantation (PPI)—remain a significant periprocedural complication.

While several individual risk factors are known (e.g., valve type, implantation depth, baseline right bundle branch block, and specific calcification locations), current preoperative planning using Computed Tomography Angiography (CTA) typically assesses these features in isolation. The collective prognostic inference of a multitude of pre-TAVR CTA features (encompassing valve leaflet calcification, aortic root dimensions, and ascending aorta geometry) on the risk of CDs is not well understood. The authors hypothesize that patient-specific anatomical variations around the aortic root, when analyzed collectively, may reveal distinct subphenotypes with differential risks for CDs.

2. Methodology

Study Cohort:

• Population: A retrospective cohort of 660 patients (330 male, 330 female) with clinically significant degenerative calcific AS undergoing TAVR evaluation.
• Sites: Stanford University and Veterans Affairs Palo Alto Health Care System (VAPAHCS).
• Data Source: Pre-TAVR CTA images.

Feature Extraction:

• Initial Extraction: 21 standard pre-TAVR CTA imaging features were extracted based on clinical guidelines.
• Feature Reduction: Highly correlated features (Pearson correlation coefficient $\ge$ 0.8) were removed to create a non-redundant set of 12 features:
  1. Calcification loads of Non-Coronary Cusp (NCC), Right Coronary Cusp (RCC), and Left Coronary Cusp (LCC).
  2. Aortic annulus area.
  3. Sinus of Valsalva (SOV) diameter at NCC.
  4. Sinotubular junction (STJ) diameter (long-axis).
  5. Ascending aorta diameter (long-axis).
  6. Right coronary artery (RCA) ostial height.
  7. Left main coronary artery (LMCA) ostial height.
  8. SOV height at NCC.
  9. SOV height at RCC.
  10. Aortic root angle.
• Covariates: Additional known risk factors (membranous septal length, LVOT calcification, implantation depth, THV type) were included in regression models.

Machine Learning Approach:

• Algorithm: Unsupervised Machine Learning (UML) using Agglomerative Hierarchical Clustering (AHC).
• Stratification: Analyses were performed separately for male and female datasets due to significant sex-specific differences in CTA features.
• Optimization:
  • Data were standardized and transformed for normality.
  • Cluster tendency was confirmed using the Hopkins statistic.
  • Ward's linkage method was selected as it yielded the highest agglomerative coefficients.
  • The optimal number of clusters was determined using 30 validated cluster indices (via the NbClust package).

Statistical Analysis:

• Outcome: A composite endpoint of TAVR-related CDs (new LBBB, transient CDs, or HAVB requiring PPI within 30 days).
• Modeling: Multivariable logistic regression assessed the association between cluster type and CDs, adjusting for conventional risk factors.
• Validation: Likelihood-ratio tests evaluated the incremental prognostic value of adding cluster type to conventional risk models.

3. Key Contributions

• Novel Application of UML: This is the first study to apply unsupervised hierarchical clustering to a comprehensive set of pre-TAVR CTA features to define patient subphenotypes specifically for CD risk stratification.
• Sex-Specific Phenotyping: The study demonstrates that anatomical clustering patterns differ significantly between sexes, necessitating separate models for accurate risk assessment.
• Beyond Single Features: It moves beyond the "additive risk" model of individual features to a "holistic phenotype" model, revealing how the combination of calcification load and root geometry (e.g., "short and wide" vs. "tall and wide") drives risk.
• Incremental Prognostic Value: The study proves that cluster assignment provides statistical improvement in model fit over traditional clinical risk factors alone.

4. Key Results

Cluster Identification:

• Male Patients: Three distinct clusters (M1, M2, M3) were identified.
  • M1 (Low Risk): Characterized by small valve leaflet calcification loads and small aortic root dimensions (both cross-sectional and height). Associated with the highest prevalence of low-gradient AS.
  • M2 (High Risk): Characterized by large calcification loads, wide aortic root (cross-sectional), but a short aortic root (height). This "short and wide" phenotype had the highest association with CDs.
  • M3 (Intermediate Risk): Characterized by large calcification loads, wide aortic root, and a tall aortic root.
• Female Patients: Two distinct clusters (F1, F2) were identified.
  • F2: Associated with larger calcification loads and larger aortic root dimensions compared to F1.
  • Differentiation: Unlike males, the female clusters did not show a significant difference in CD risk between them.

Risk Association (Logistic Regression):

• Male Risk: Compared to the reference M1, M2 was significantly associated with higher CD risk (Odds Ratio [OR] = 2.15, $P=0.032$). M3 showed a trend but was not statistically significant compared to M1 (OR = 2.12, $P=0.085$). There was no significant difference between M2 and M3.
• Female Risk: No significant difference in CD risk was found between F1 and F2 (OR = 1.29, $P=0.581$).
• Model Improvement: Including cluster type in a multivariable regression model significantly improved the goodness-of-fit ($P=0.020$) compared to a model with only conventional risk factors.

Clinical Characteristics:

• The clusters were not distinguishable by basic demographics or comorbidities but were clearly distinguishable by CTA anatomical features.
• M2 patients had significantly smaller RCA/LMCA ostial heights and SOV heights compared to M3, suggesting a "short and wide" anatomical constraint.

5. Significance and Conclusion

• Personalized Risk Stratification: The study suggests that TAVR planning should move from analyzing isolated measurements to identifying specific anatomical subphenotypes. Specifically, male patients with a "short and wide" aortic root combined with heavy calcification (M2) are at a uniquely elevated risk for conduction disturbances.
• Clinical Workflow Integration: These findings advocate for the integration of UML algorithms into pre-TAVR CTA software. By automatically extracting features and assigning patients to risk clusters, physicians could better anticipate CD risks.
• Mitigation Strategies: Identifying high-risk phenotypes (like M2) could guide procedural strategies, such as selecting specific Transcatheter Heart Valve (THV) types, adjusting implantation depth, or planning for closer post-procedural monitoring.
• Future Directions: While the female clusters did not show differential CD risks in this cohort (potentially due to lower overall event rates), the methodology establishes a framework for refining sex-specific risk models with larger datasets.

In summary, this paper demonstrates that unsupervised machine learning of pre-TAVR CTAs can uncover hidden anatomical subphenotypes that significantly improve the prediction of conduction disturbances, particularly in male patients, offering a pathway toward more personalized and safer TAVR procedures.