📊 statistics

Calibrated Alzheimer's Conversion Risk in Mild Cognitive Impairment: Persistent Homology of Clinical Trajectories with Conformal Guarantees

This study introduces a leakage-audited, conformally calibrated machine learning pipeline that utilizes persistent homology of clinical trajectories to predict Alzheimer's conversion in mild cognitive impairment with improved accuracy, topological biomarker discovery, and rigorous individual-level uncertainty guarantees.

Original authors: Navin Bondade

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Navin Bondade

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a long, winding road trip. Some cars drive smoothly, others sputter, and a few take sudden, sharp turns. In the world of medicine, doctors often try to predict which cars (patients) will eventually break down completely (develop Alzheimer's disease) based on how they are driving right now (their current symptoms). This is a tricky job because the road is full of fog, and the data can be messy. Scientists use a field called machine learning to build "crystal balls" that predict these breakdowns. But here's the catch: sometimes these crystal balls are too confident, or they cheat by peeking at the future to make their predictions look better than they really are. To fix this, researchers are now using a branch of math called Topological Data Analysis (TDA). Think of TDA not as looking at the numbers on the dashboard, but as studying the shape of the entire journey. It asks: Is the path a straight line? Is it a jagged zig-zag? Does it form a loop? By measuring the "shape" of a patient's health history over time, scientists hope to spot the early signs of a crash that simple numbers might miss.

This paper is about a team of researchers who built a new, super-strict crystal ball to predict if a person with Mild Cognitive Impairment (MCI)—a state of mild memory trouble—will turn into full-blown Alzheimer's. They didn't just build a better predictor; they built a "leak-proof" one. They found that many previous studies accidentally cheated by letting future information sneak into their training data, making their models look like geniuses when they were actually just lucky. The authors fixed these leaks and introduced a new way to measure the "shape" of a patient's decline using something called persistent homology. Imagine taking a patient's test scores and brain scans over several years and turning them into a cloud of dots. Persistent homology is like blowing bubbles around those dots to see how they connect. The researchers discovered that the "shape" of this cloud (specifically, a measure called H0 persistence entropy) is a powerful clue: people who are destined to convert to Alzheimer's tend to have a very rigid, predictable, and monotone decline, while those who stay stable have a more chaotic, irregular path.

The team tested their new system on 741 people from a large database called ADNI. They were very careful to avoid "data leakage," which is like a student peeking at the answer key before taking a test. They found that if you don't fix these leaks, your model might look like it's 93.4% accurate, but once you clean up the cheating, the real accuracy drops to about 85.9%. That 7.5% difference is huge in science; it means previous studies were likely overestimating how well they could predict the future. Their new model, which combines these "shape" measurements with standard medical data, achieved an accuracy of about 86% to 88% on new, unseen data. More importantly, they added a safety feature called conformal prediction. Instead of just saying "You have a 70% chance of getting sick," the model says, "We are 90% sure you are either in the 'safe' group or the 'at-risk' group." If the model is truly unsure, it admits it by giving a "maybe" answer, rather than guessing blindly. This is a first for this type of medical prediction.

The researchers also found a fascinating link between this "shape" of decline and genetics. They discovered that the "entropy" (or messiness) of the health trajectory is lower in people who carry the APOE4 gene, a known risk factor for Alzheimer's. In simple terms, people with this gene tend to decline in a very straight, boring line, whereas those without it might have ups and downs. This suggests that the "shape" of the decline is biologically real and not just a random math trick. However, the paper is careful to note that while this "shape" feature is the most important clue in their model, it doesn't make the model perfect. The model is still better at predicting when someone might get sick (survival analysis) than just if they will get sick in a fixed four-year window.

One of the most exciting parts of this work is how honest it is about its own limits. The model works well on the specific group of people it was tested on, but the researchers warn that if you try to use it on a completely different group of people (like a different country or a different hospital), the confidence scores might be too high. They showed that the model needs to be "recalibrated" before it can be used in the real world, much like a compass needs to be adjusted when you move from one side of the planet to the other. They also checked if the model was fair to different groups of people based on race, gender, and age. They found that the model was surprisingly fair, with very small differences in how often it made mistakes for different groups, which is a big deal because many medical AI tools have been unfair in the past.

In the end, this paper doesn't claim to have "solved" Alzheimer's prediction. Instead, it offers a new, more honest toolkit. It gives us a way to look at the shape of a patient's journey, a method to catch cheating in scientific studies, and a way to tell patients, "We are this sure, and here is where we are unsure." It suggests that the future of medical prediction isn't just about having more data, but about having cleaner data, better math to understand the shape of that data, and the humility to admit when we don't know the answer yet. The authors propose that their "shape" measurement (H0 persistence entropy) is a new kind of biological marker that deserves more study, potentially helping doctors spot the early, subtle signs of Alzheimer's before it's too late to plan for care.

Technical Summary: Calibrated Alzheimer's Conversion Risk in Mild Cognitive Impairment

Problem Statement
Predicting the conversion from Mild Cognitive Impairment (MCI) to Alzheimer's Disease (AD) is critical for clinical care planning and trial enrichment. However, existing machine learning models for this task suffer from three primary deficiencies: (1) a lack of individual-level uncertainty estimates, (2) a frequent absence of transparent data leakage audits, and (3) limited reporting on algorithmic fairness across demographic subgroups. Furthermore, longitudinal Electronic Health Record (EHR) derived cohorts are prone to non-obvious data leakage sources that artificially inflate performance metrics.

Methodology
The study utilizes data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), analyzing 741 MCI subjects (240 converters, 32.4%) with a uniform 4-year follow-up cap applied to both converters and non-converters to ensure consistent observation windows.

Data Leakage Audit and Correction: The authors identified and corrected five specific sources of data leakage:
1. Post-conversion visit contamination.
2. Informative censoring due to asymmetric follow-up.
3. Cross-cohort subject overlap in external validation.
4. Global-scaler contamination in within-cross-validation (CV) feature extraction.
5. Non-exchangeable splits in conformal calibration.
  A "naïve" pipeline without these corrections achieved an AUC of 0.934, which was inflated by +0.075 compared to the corrected value.
Feature Extraction and Topological Data Analysis (TDA):
- Trajectory Data: Twelve longitudinal biomarkers (including MMSE, ADAS-13, CDRSB, and volumetric measures) were treated as point clouds in $\mathbb{R}^{12}$ .
- Persistent Homology: Vietoris-Rips persistent homology was applied to these point clouds.
  - $H_0$ (Connected Components): Summarized by six statistics, including persistence entropy, derived directly from the clinical trajectory point clouds.
  - $H_1$ (Loops): Computed via Takens delay embedding on MMSE, ADAS-13, and hippocampal volume.
- Additional Features: The TDA features were combined with trajectory slopes, acceleration terms, sublevel-set proxies, and 14 baseline/demographic variables, resulting in a 76-variable feature set.
- Normalization: Features were normalized within CV folds to prevent leakage.
Modeling and Evaluation:
- Ensemble: A stacking ensemble (Stack) combining Random Forest (RF), Gradient Boosting (GBM), XGBoost, and SVM with a logistic regression meta-learner.
- Validation: Evaluated via 5-fold stratified cross-validation with 1,000-resample bootstrap confidence intervals. A nested feature selection protocol was used to quantify optimism from SHAP-based feature selection.
- External Validation: Models trained on ADNI-1 (2004–2005) were applied without retraining to ADNI-2/GO/3 (2010–2021), ensuring zero subject overlap and a temporal external validation with a 32.8-percentage-point prevalence shift.
- Conformal Prediction: Split-conformal prediction with class-conditional (Mondrian) quantiles was applied to provide distribution-free individual risk guarantees.

Key Results

Performance Metrics:
- The corrected pipeline achieved an internal AUC of 0.859 (Stack+Top20: 0.866).
- On the external ADNI-2/GO/3 cohort, the pipeline achieved an AUC of 0.879.
- The model outperformed a CDRSB-only baseline (+0.201 AUC) and a three-feature clinical baseline (MMSE, hippocampus, APOE4; +0.110 AUC).
- TDA Contribution: In fixed-horizon classification, the marginal contribution of TDA features was small and non-significant ( $\Delta$ AUC = +0.006, $p=0.117$ ). However, in survival analysis, TDA features provided clear improvements: Penalised Cox concordance increased from 0.753 to 0.799 (+0.045), and Random Survival Forest concordance increased from 0.812 to 0.826 (+0.014).
Topological Biomarker Validation:
- $H_0$ Persistence Entropy was the single most important predictor (SHAP rank 1), selected in 90% of bootstrap resamples.
- It showed a significant negative association with APOE4 allele dosage (Spearman $r = -0.191$ , $p < 0.0001$ ), suggesting that APOE4 carriers exhibit more stereotyped, monotone clinical decline (lower entropy). This association persisted after controlling for disease stage and visit count, though it was partially confounded by observation density.
Uncertainty Quantification and Fairness:
- Conformal Coverage: The pipeline achieved 93.6% internal coverage (target 90%) and 96.9% external empirical coverage under the prevalence shift. Class-conditional coverage met targets for both converters (91.8%) and non-converters (98.8%).
- Fairness: The maximum false-negative rate gap across seven demographic subgroups was 0.092, significantly lower than the 0.28 reported in prior work. No statistically significant performance disparities were found across subgroups.
Calibration: External calibration slopes indicated overconfident probability estimates (slope = 2.06). Platt scaling on a held-out external subset corrected this (slope = 0.681), making the probabilities suitable for clinical use after recalibration.

Significance and Claims
The paper claims three primary contributions:

Leakage Audit Framework: It provides a replicable framework identifying five sources of data leakage in longitudinal ADNI studies, demonstrating that uncorrected pipelines can inflate AUC by +0.075.
Topological Biomarker: It proposes $H_0$ persistence entropy as a biologically motivated topological biomarker of cognitive decline, validated by its association with APOE4 dosage and its superior performance in survival modeling compared to fixed-horizon classification.
Individual-Level Risk Guarantees: It introduces the first split-conformal individual risk guarantees for MCI-to-AD conversion, providing formally valid, distribution-free uncertainty quantification (prediction sets) that prior models lacked.

The authors modestly state that while the discrimination performance is competitive with tabular benchmarks, the primary contribution is methodological rigour and individual-level uncertainty quantification. They emphasize that TDA's predictive signal is most naturally expressed in survival frameworks rather than fixed-horizon classification and that the pipeline is not a state-of-the-art discriminator compared to multimodal deep learning models using MRI, but rather a robust, transparent, and calibrated approach for tabular clinical data.

Technical Summary: Calibrated Alzheimer's Conversion Risk in Mild Cognitive Impairment

More like this