New version has been released. Please make sure to use latest version.
PANDORA
GithubaTomic LabPANDORA Website

CardioBoost Workflow

Phase 1: Data Configuration & Initial Exploration

Purpose: Upload the dataset, inspect its structure, handle missing data (if any), and perform initial exploratory data analysis to understand data distributions and basic relationships between features, treatment, and the cardio outcome.

Actions:

  1. Launch PANDORA & Upload Data:

    • Start your PANDORA instance.

    • Navigate to Workspace.

    • Upload your cardioguard_trial_data.csv dataset (ensure it includes the hypothetical TreatmentGroup column: e.g., Placebo=0, CardioGuard=1).

    • Select the uploaded dataset for analysis.

  2. Initial Data Inspection (Data Overview):

    • Navigate to Discovery -> Start -> Data Overview.

    • Select key columns for initial review: TreatmentGroup, Age, ap_hi, ap_lo, cholesterol, gluc, smoke, active, and cardio.

    • Examine the Distribution Plot and Table Plot.

      • Check data types. Note that cholesterol and gluc are categorical. Age is in days; consider if transformation to years is needed outside PANDORA for easier interpretation, though models can handle raw days.

      • Assess distributions for variables like ap_hi, ap_lo.

      • Note any missing values (NAs).

    • In the Side Panel, under Preprocessing:

      • If NAs are present: medianImpute for numerical features (like Weight, ap_hi, ap_lo if they have NAs) or knnImpute. For categorical features with NAs (less common for cholesterol or gluc if coded 1/2/3, but possible), PANDORA's imputation might convert them or specific handling might be needed (e.g., imputing with mode).

      • Apply center and scale to numerical predictors, as this is generally good practice for many PANDORA models.

      • zv (zero variance) and nzv (near-zero variance) can be used.

    (Conceptual PANDORA Screenshot)

    [PANDORA Interface: Discovery -> Data Overview showing columns like 'ap_hi', 'cholesterol', 'TreatmentGroup', 'cardio', with distribution plots. Categorical nature of 'cholesterol' and 'gluc' visible.]

  3. Explore Correlations:

    • Navigate to Discovery -> Correlation.

    • Column Selection: Select TreatmentGroup (coded numerically), Age, Height, Weight, ap_hi, ap_lo, numerically coded versions of cholesterol and gluc (or understand how PANDORA handles categoricals here – it typically expects numeric input for standard correlation matrices), smoke, alco, active, and cardio.

    • Correlation Method: Spearman can be a good choice given mixed data types and potential non-linearities.

    • Correlation Settings:

      • NA Action: pairwise.complete.obs.

      • Plot Method: circle or number.

    • Significance: Enable significance testing and p-value adjustment (e.g., BH).

    • Click Plot Image.

    • Interpretation:

      • Examine the correlation between TreatmentGroup and cardio. A negative correlation would be hoped for if CardioGuard is effective (less cardio=1 in the CardioGuard group).

      • Look at correlations between risk factors (e.g., ap_hi, cholesterol, smoke) and cardio.

    (Conceptual PANDORA Screenshot )

    [PANDORA Interface: Discovery -> Correlation showing a correlogram. A circle/number indicating the correlation between a numerically coded 'TreatmentGroup' and 'cardio' would be of key interest, alongside known risk factors and 'cardio'.]

Summary of Phase 1: The dataset is uploaded and initially explored. Basic cleaning is considered. Preliminary correlations might offer early hints about the drug's association with the cardio outcome and highlight relationships between risk factors.

Phase 2: Understanding the Outcome Variable (`cardio`)

Purpose: The primary outcome variable cardio (Presence or absence of cardiovascular disease) is already defined and binary in your dataset. This phase focuses on understanding its prevalence and relationship with key baseline features.

Actions:

  1. Assess Prevalence of cardio:

    • In Discovery -> Data Overview, select the cardio column.

    • The table plot or distribution plot (for a binary variable, it will show counts/proportions) will indicate the number of individuals with (cardio=1) and without (cardio=0) cardiovascular disease. This is important for understanding class balance for the subsequent classification modeling.

  2. Initial Stratification (using PCA/t-SNE for visualization):

    • Navigate to Discovery -> Start -> PCA Analysis.

    • Column Selection: Select all baseline predictor variables (Objective, Examination, Subjective features, excluding PatientID and TreatmentGroup from the PCA calculation itself).

    • Preprocessing: Ensure center and scale are applied to numerical features. Handle NAs. PANDORA should manage categorical features like cholesterol, gluc, Gender appropriately for PCA (often via MCA for categorical variables or dummy coding if numerical PCA is forced).

    • PCA Settings:

      • Grouping Variable: Select cardio.

      • Observe if individuals with and without cardiovascular disease form distinct clusters or overlap based on their baseline characteristics.

    • Click Plot Image.

    • Interpretation: This provides a visual sense of whether baseline profiles inherently separate those with and without the cardio outcome.

    (Conceptual PANDORA Screenshot for PCA)

    [PANDORA Interface: Discovery -> PCA Analysis -> Individuals Plot, with points colored by 'cardio' status. This helps visualize if baseline features differentiate those with vs. without cardiovascular disease.]

Summary of Phase 2: The prevalence of the cardio outcome is assessed, and initial visualizations explore how baseline characteristics relate to this outcome.

Phase 3: Confounding Variable Check

Purpose: Assess whether key baseline characteristics (e.g., Age, Gender, Weight, ap_hi, cholesterol) are evenly distributed across the TreatmentGroup (CardioGuard vs. Placebo). Significant imbalances in a randomized trial would be concerning.

Actions:

  1. Visualize Distributions by Treatment Group (using PCA or t-SNE):

    • Navigate to Discovery -> Start -> PCA Analysis (or t-SNE Analysis).

    • Column Selection: Select all relevant baseline predictor variables (Objective, Examination, Subjective features).

    • Preprocessing: Ensure center and scale are applied to numerical features. Handle NAs.

    • PCA Settings / t-SNE Settings:

      • Grouping Variable: Select TreatmentGroup.

      • Observe if the CardioGuard and Placebo arms form distinct clusters or overlap significantly based on baseline characteristics. In a well-randomized trial, they should largely overlap.

    • Click Plot Image.

  2. Compare Individual Confounders (Manual review or using Correlation if applicable):

    • Data Overview: While PANDORA's Data Overview might not directly provide grouped summary statistics (e.g., mean age for placebo vs. mean age for CardioGuard), you can select individual confounders and visually inspect distributions.

    • Correlation (from Phase 1): Check the correlation matrix. Was there any strong, unexpected correlation between a baseline feature like Age and the numerically coded TreatmentGroup? This would be an issue.

    (Conceptual PANDORA Screenshot for PCA)

    [PANDORA Interface: Discovery -> PCA Analysis -> Individuals Plot, with points colored by 'TreatmentGroup'. Ideally, Placebo and CardioGuard groups should largely overlap, indicating good baseline balance from randomization.]

Summary of Phase 3: Potential confounding variables are checked for imbalances across treatment groups. This step is crucial for validating the randomization process.

Phase 4: Predictive Modeling for Cardiovascular Disease (`cardio`)

Purpose: Build a classification model to predict the cardio outcome (presence/absence of cardiovascular disease). The key goal is to determine if TreatmentGroup (CardioGuard vs. Placebo) is an important predictor.

Actions (using PANDORA Predictive - SIMON):

  1. Navigate to Predictive Modeling:

    • Select your dataset in Workspace.

    • Navigate to Predictive -> Start.

  2. Setup for CLASSIFICATION Model (Predicting cardio):

    • Analysis Properties:

      • Select Classification as the analysis type.

      • Predictor Variables:

        • Include TreatmentGroup (ensure numerically coded: e.g., Placebo=0, CardioGuard=1).

        • Include all other baseline Objective, Examination, and Subjective features (Age, Height, Weight, Gender, ap_hi, ap_lo, cholesterol, gluc, smoke, alco, active).

      • Response: Select cardio.

      • Training/Testing Dataset Partition (%): e.g., 75% for training, 25% for testing.

      • Preprocessing: Apply center, scale (for numerical predictors), appropriate NA handling (e.g., medianImpute or knnImpute), zv, nzv. PANDORA's models or preprocessing steps should handle the categorical nature of cholesterol, gluc, and Gender (e.g., through internal dummy coding).

    • Model Selection and Customization:

      • Select models suitable for binary classification (e.g., glm for logistic regression, rf - Random Forest Classifier, C5.0, AdaBoost.M1, mlp - Neural Network Classifier).

    • Advanced Options:

      • Consider enabling Feature Selection. This will help identify which variables (hopefully including TreatmentGroup) are most important for predicting cardio.

    • Click Validate data, then Process. Name this task appropriately (e.g., "CardioGuard_Cardio_Prediction").

  3. Monitor Progress: Check the Dashboard for model training completion.

    (Conceptual PANDORA Screenshot for SIMON Setup)

    [PANDORA Interface: Predictive -> Start (SIMON) showing 'Classification' selected. 'cardio' as Response. 'TreatmentGroup', 'Age', 'ap_hi', 'cholesterol', 'smoke', etc., as Predictors. Preprocessing options checked.]

Summary of Phase 4: A classification model is configured and run to predict cardiovascular disease status. The TreatmentGroup variable is included as a key predictor to assess the drug's potential impact.

Phase 5: Analyzing Predictive Results & Drug Effect

Purpose: Evaluate the performance of the trained classification models and, most importantly, interpret the role and importance of the TreatmentGroup variable in predicting cardio status.

Actions (using PANDORA Predictive - Exploration):

  1. Select Queue for Exploration:

    • Navigate to the Dashboard. Select your completed "CardioGuard_Cardio_Prediction" task.

    • Navigate to Predictive -> Exploration.

    • Configure Exploration Space: Select cardio as the response outcome, relevant metrics (e.g., PredictAUC, TrainAUC, BalancedAccuracy, F1-Score, Precision, Recall), the dataset, and the models to evaluate.

  2. Evaluate Model Performance:

    • Examine the metrics table for the test set.

    • ROC Curve Analysis: Compare Training ROC and Testing ROC curves. A higher Test AUC indicates better model discrimination for predicting cardio status. Check for overfitting.

    • Training Summary: Compare different models if multiple were run.

  3. Assess Drug Effectiveness (Importance of TreatmentGroup):

    • Variable Importance Tab:

      • Select your best performing model(s).

      • Examine the bar plot of feature importance. Is TreatmentGroup a significant predictor of cardio? A high rank and importance score would suggest CardioGuard influences the likelihood of having cardiovascular disease.

      • (Conceptual Interpretation: If CardioGuard reduces risk, and 'TreatmentGroup' (CardioGuard=1) is negatively associated with 'cardio'=1, it should appear as an important feature.)

      • Use the "Features across dataset" sub-tab to select TreatmentGroup and other top predictors. Visualize how their values differ between individuals with cardio=0 and cardio=1.

    (Conceptual PANDORA Screenshot for Variable Importance)

    [PANDORA Interface: Predictive -> Exploration -> Variable Importance, showing a bar chart where 'TreatmentGroup' has a notable importance score for predicting 'cardio'.]

    • Model Interpretation Tab (Explainable AI - xAI):

      • Select your best model.

      • For TreatmentGroup:

        • PDP (Partial Dependence Plots) / ICE (Individual Conditional Expectation) Plots: Plot the predicted probability of cardio=1 based on the TreatmentGroup variable. This will visually show the average effect of CardioGuard vs. Placebo on the likelihood of having cardiovascular disease.

        • (Conceptual Interpretation: An ICE plot for 'TreatmentGroup' might show that when TreatmentGroup changes from 0 (Placebo) to 1 (CardioGuard), the predicted probability of 'cardio'=1 decreases, indicating a protective effect.)

        • LIME Plot: Explain predictions for individual hypothetical patients to see how TreatmentGroup contributed to their predicted cardio status.

    (Conceptual PANDORA Screenshot for PDP/ICE Plot)

    [PANDORA Interface: Predictive -> Exploration -> Model Interpretation, showing a PDP/ICE plot for 'TreatmentGroup'. The y-axis shows predicted probability of 'cardio'=1. A lower probability for the CardioGuard group would suggest a positive drug effect.]

Summary of Phase 5: Model performance is assessed. The importance and effect of TreatmentGroup are investigated to understand if and how CardioGuard influences the cardio outcome.

Phase 6: Synthesizing Findings

Purpose: Combine all analyses to report on the potential effectiveness of CardioGuard in relation to cardiovascular disease status, the performance of predictive models, and insights into relevant risk factors.

Actions & Report Structure (Conceptual):

  1. Overall Drug Effect on cardio Outcome:

    • Summarize the findings on the importance of TreatmentGroup from the classification models.

    • Report the effect observed from PDP/ICE plots (e.g., "CardioGuard was associated with an X% reduction in the predicted probability of having cardiovascular disease (cardio=1) compared to Placebo.").

  2. Best Predictive Model(s) for cardio:

    • Report the best-performing model based on test set metrics (e.g., Test AUC, Balanced Accuracy).

    • Include key plots: ROC curves, Variable importance plots (highlighting TreatmentGroup and other top predictors like ap_hi, cholesterol, smoke, etc.).

  3. Key Predictors of Cardiovascular Disease (Risk Factors):

    • List other baseline variables that were consistently important in predicting cardio (e.g., Age, ap_hi, cholesterol, smoke, active).

    • Discuss insights from "Features across dataset" plots or Model Interpretation plots regarding these risk factors.

  4. Confounding Assessment:

    • Briefly mention the results of the confounding checks (Phase 3).

  5. Limitations:

    • Acknowledge limitations (e.g., hypothetical nature of data, specific features available, cross-sectional cardio outcome rather than incident disease over time if that were the true trial design).

  6. Conclusion & Future Directions:

    • Conclude on the evidence for CardioGuard's potential effect on cardiovascular disease status based on the PANDORA analysis.

    • Suggest if the predictive models are useful for risk stratification.

Example Conclusion: PANDORA analysis of the hypothetical CardioGuard trial suggests that treatment with CardioGuard is an important factor associated with a reduced likelihood of having cardiovascular disease (best model Test AUC: 0.75). The TreatmentGroup variable was a key predictor, and PDP/ICE plots indicated an estimated 15% absolute risk reduction in predicted cardio status for the CardioGuard group. Other significant predictors included ap_hi, cholesterol (level 3 vs 1), and smoke. These findings warrant further investigation into CardioGuard as a cardiovascular protective agent.

PreviousDataset Description

Last updated

Was this helpful?