📄 medicine

Development and Validation of a Nomogram-Based Clinical Prediction Model for Congestive Heart Failure in Patients with Parkinson's Disease

This study developed and internally validated a nomogram-based clinical prediction model using 11 routinely available variables from the MIMIC-IV database to effectively identify and assess the risk of congestive heart failure in patients with Parkinson's disease.

Original authors: Dan Yang, Xin Xu, Shizao Fei, Changqing Zhan

Published 2026-07-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dan Yang, Xin Xu, Shizao Fei, Changqing Zhan

Original paper licensed under CC BY 4.0 (https://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 a doctor treating a patient with Parkinson's disease. You know that Parkinson's affects movement, but this study suggests it also quietly stresses the heart, sometimes leading to a condition called Congestive Heart Failure (CHF). The problem is that the symptoms of a failing heart (like tiredness or shortness of breath) can look exactly like the symptoms of Parkinson's, making it hard to tell what is causing what until it's too late.

The researchers in this paper wanted to build a "early warning system" to spot patients at risk before things get serious. Here is how they did it, explained simply:

1. The Data: A Digital Time Capsule

The team didn't run new experiments on patients. Instead, they went into a massive digital library called MIMIC-IV. Think of this library as a giant warehouse containing the medical records of thousands of people who were treated in a Boston hospital between 2008 and 2019. They pulled out the records of 293 people who had Parkinson's disease.

2. The Detective Work: Finding the Clues

They had a mountain of data for each patient: blood tests, heart rates, breathing rates, age, and more. They needed to find the specific clues that signaled a heart failure risk.

The First Filter (LASSO): Imagine they had a huge bag of mixed-up puzzle pieces. They used a computer method called LASSO regression to quickly sort through the bag and pull out only the pieces that actually fit the picture. This narrowed their list down from dozens of possibilities to the most important 16 clues.
The Second Filter (Logistic Regression): They then used a second method to fine-tune the list. They asked, "If we remove this clue, does the picture get worse?" They ended up with a final list of 11 specific variables that worked best together.

3. The 11 Clues (The Ingredients)

The final "recipe" for predicting heart failure risk included a mix of things you might expect and some surprises:

Age: Older patients were at higher risk.
Kidney Health: Lower minimum creatinine levels (a kidney marker) were linked to higher risk.
Heart & Lungs: Faster maximum breathing rates and slower minimum heart rates were warning signs.
Blood Chemistry: High maximum levels of a heart enzyme (CK-MB), high bicarbonate, high blood sugar, and high LDH were risk factors.
Blood Coagulation: Specific measures of how fast blood clots (INR and PTT) played a role.
The Good News: Higher maximum hemoglobin (a measure of healthy blood cells) actually protected against the risk.

4. The Tool: A "Risk Score" Calculator

To make this useful for doctors, they built a Nomogram.

The Analogy: Think of a Nomogram like a sliding ruler or a customized calculator. Instead of a doctor doing complex math in their head, they just draw a line across a chart.
How it works: You take the patient's 11 numbers (age, blood sugar, etc.), find them on the chart, and draw a line down to the bottom. The spot where the line lands tells you the percentage chance that this specific patient will develop heart failure.

5. Did It Work?

The researchers tested their new tool to see if it was accurate:

Discrimination (The "Taste Test"): The tool scored 0.832 on a scale where 1.0 is perfect. This is like a chef who can correctly identify the difference between salt and sugar 83% of the time. It's considered "good."
Calibration (The "Weather Forecast"): If the tool says a patient has a 20% risk, do 20 out of 100 similar patients actually get sick? The answer was yes; the predictions matched reality very well.
Comparison: They tried making a simpler version with only 8 clues, but the 11-clue version was significantly better at spotting the risk.

6. The Catch (Limitations)

The authors are very honest about the tool's current limits:

Small Sample Size: They only looked at 293 people. It's like testing a new car on a short track; it might work there, but we don't know how it handles a long highway yet.
Single Location: All the data came from one hospital in Boston. The tool might need to be tested on people in different cities or countries to ensure it works everywhere.
Retrospective: They looked at past records, not future events.

The Bottom Line

This paper presents a new, practical calculator (the Nomogram) that uses 11 routine medical checks to estimate how likely a Parkinson's patient is to develop heart failure. It works well on the data it was tested with, but the authors say it needs to be tested on a larger, more diverse group of people before doctors should start using it in real clinics.

Technical Summary: Development and Validation of a Nomogram-Based Clinical Prediction Model for Congestive Heart Failure in Patients with Parkinson's Disease

Problem Statement
Patients with Parkinson's disease (PD) face a significantly elevated risk of cardiovascular complications, particularly congestive heart failure (CHF). Despite the known systemic effects of PD on the cardiovascular system—including cardiac autonomic dysfunction, left ventricular hypertrophy, and diastolic dysfunction—validated clinical prediction tools specifically designed to identify CHF risk within the PD population are currently lacking. The overlap in symptoms between PD and CHF (e.g., fatigue, reduced exercise tolerance) complicates diagnosis and management. Consequently, there is an urgent need for early identification tools to facilitate timely intervention and improve clinical outcomes for this high-risk cohort.

Methodology
This retrospective study utilized data from the MIMIC-IV database (Medical Information Mart for Intensive Care), covering patients admitted to Beth Israel Deaconess Medical Center between 2008 and 2019.

Study Population: The final analysis included 293 patients diagnosed with Parkinson's disease (identified via ICD-10 code G20). Of these, 110 (37.5%) were classified as having CHF (ICD-10 code I50.x), and 183 served as the non-CHF control group. For patients with multiple admissions, only the first eligible record was used.
Variable Selection: A two-step statistical approach was employed to construct the model:
1. LASSO Regression: Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation was used to screen a broad pool of routinely available clinical variables, reducing dimensionality and addressing multicollinearity. This yielded 16 candidate variables.
2. Backward Stepwise Logistic Regression: The 16 candidates were further refined using backward stepwise multivariable logistic regression guided by the Akaike Information Criterion (AIC) to determine the final set of predictors.
Model Construction: A nomogram was developed based on the final logistic regression model to visualize individual risk probabilities.
Validation and Evaluation: Internal validation was performed using bootstrap resampling. Model performance was assessed via:
- Discrimination: Area under the receiver operating characteristic curve (AUC).
- Calibration: Calibration curves comparing predicted vs. observed probabilities.
- Clinical Utility: Decision Curve Analysis (DCA).
- Statistical Comparison: The DeLong test was used to compare the AUC of the final model against simplified versions (e.g., an 8-variable model).

Key Contributions
The study developed and validated an 11-variable clinical prediction model and a corresponding nomogram specifically for CHF risk in PD patients. The model integrates routinely available clinical variables, avoiding the need for specialized software or complex calculations.

The final model includes the following 11 predictors:

Age
Minimum Creatinine
Maximum Respiratory Rate
Minimum Heart Rate
Maximum CK-MB (Creatine Kinase-MB)
Maximum Bicarbonate
Minimum Blood Glucose
Maximum Hemoglobin
Minimum INR (International Normalized Ratio)
Minimum PTT (Partial Thromboplastin Time)
Maximum LDH (Lactate Dehydrogenase)

Notably, while Minimum INR, Minimum PTT, and Maximum LDH did not reach statistical significance ( $P < 0.05$ ) individually, they were retained in the final model because their inclusion improved the overall model fit (lowered AIC) according to the backward stepwise algorithm. Maximum Hemoglobin was identified as a protective factor.

Results

Performance: The final 11-variable model achieved an AUC of 0.832, indicating good discriminative ability. The optimal cutoff value was determined to be 0.366, with a sensitivity of 0.773 and specificity of 0.760.
Comparison: The 11-variable model demonstrated significantly better discriminative performance than a simplified 8-variable model (derived by removing non-significant variables), with a DeLong test result of $Z = 2.5305$ ( $P = 0.01139$ ). While a 12-variable model showed a marginally higher AUC (0.838), the 11-variable model was selected for its optimal balance of performance and parsimony (AIC = 305.492).
Calibration and Utility: Calibration curves showed satisfactory agreement between predicted and observed probabilities. Decision Curve Analysis (DCA) indicated that the model provides a favorable net clinical benefit across a threshold probability range of approximately 0.05 to 0.45.
Stability: Multicollinearity diagnostics confirmed that all Variance Inflation Factor (VIF) values were below 2, supporting the stability of the model structure.

Significance and Claims
The authors claim that the developed nomogram offers a practical, accessible tool for the early identification and individualized risk assessment of CHF in patients with Parkinson's disease. By leveraging routinely available clinical variables, the model addresses a gap in current clinical practice where validated prediction tools for this specific comorbidity are absent.

The study emphasizes that the model demonstrates good discrimination, calibration, and clinical utility, potentially aiding clinicians in stratifying risk and allocating medical resources more rationally. However, the authors maintain a modest stance regarding the model's immediate deployment, explicitly stating that external validation using larger, multicenter cohorts is required before the model can be recommended for broader clinical application. They acknowledge limitations including the relatively small sample size (293 patients), the single-center nature of the data (MIMIC-IV), and the potential for residual confounding inherent in retrospective observational studies.