CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support

arXiv:2605.20468 · cs.LG, stat.ME, stat.ML · Submitted 2026-05-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "CASCADE Conformal Prediction".

Tom: Effective medication management in Parkinson’s Disease (PD) is challenging due to heterogeneous disease progression, variable patient response, and medication side effects.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So we’ve talked about the core mechanism of this paper, and now let’s look at the title and who wrote this piece, "CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support." It really tells you exactly what it's aiming to do.

Jane: It sounds technical, but if we break it down simply, they are proposing a new method for making prediction intervals that adjust based on the uncertainty of the decision process itself.

Lu: The authors include Ricardo Diaz-Rincon and Muxuan Liang, and their work is rooted in bridging classification tasks with regression tasks through this uncertainty propagation.

Meng: I see what they mean; it’s about taking a binary task—like needing a medication change—and using its uncertainty to inform the continuous prediction of how much dosage adjustment is needed.

Lalam: The title really highlights that the key innovation here isn't just predicting the LEDD, but making that prediction interval adaptive based on whether we are sure about what needs to happen first.

Tom: Right, so instead of a fixed safety margin in the prediction interval, they’re proposing something that actively grows or shrinks depending on how uncertain we are at each stage.

Jane: That’s the simple idea: if the initial decision is very clear, you get a tight interval; if it's ambiguous, you get a wider one that warns the clinician about higher risk.

Lu: This framework is essentially creating a mathematical bridge between discrete clinical decisions and continuous dose forecasting by dynamically scaling prediction intervals based on upstream classification reliability.

Meng: It sounds like they’re trying to solve the problem where we have models that give us a number, but we don't know how much to trust that number when the underlying situation is messy.

Lalam: That bridge between discrete decision-making and continuous forecasting is what makes this paper so relevant because it directly addresses the difficulty in managing heterogeneous disease progression in Parkinson’s Disease.

Tom: Exactly; they are taking a complex clinical problem and applying a sophisticated probabilistic framework to get more reliable, context-aware dosing recommendations.

Jane: It shows that we don't have to treat every prediction with the same level of caution just because the model produced an interval around it.

The paper's summary: Tom: We’ve covered the basics of what CASCADE is, and now let’s get into the actual summary of what this paper describes in "CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support."

Jane: The summary explains that they introduce a novel conformal prediction framework specifically designed to propagate epistemic uncertainty from a screening classifier to adapt downstream predictions.

Lu: They describe the two-stage architecture where Stage one is the intervention assessment—determining if a change is needed or if the patient's state is stable—and Stage two is the dosage prediction task.

Meng: In simple terms, they are using that first stage to get a clear signal about whether we need to adjust medication before even worrying about how much we should adjust it.

Lalam: The core concept involves extracting the Venn-Abers Uncertainty Score, uVA(x) directly from the Stage one predicted probability pˆ(x), which measures the length of that multi-probabilistic interval.

Tom: That uVA(x) then flows into a scaling function sigma(x), which is defined by a sensitivity parameter beta, allowing it to dynamically scale the prediction interval based on this upstream uncertainty.

Jane: So, the ultimate goal is to create an adaptive prediction interval Cˆ(x) that expands or shrinks depending on the patient’s level of ambiguity identified in Stage one.

Lu: They detail two ways they achieve this: one is a discrete approach called Mondrian, and another is a continuous approach based on Normalized Conformal Prediction.

Meng: The continuous version uses a scaling function σ(x) = one + beta

uVA(x)/u¯VA − one: , which parameterizes the sensitivity by that factor beta, where u¯VA is the mean uncertainty across the calibration set.

Lalam: This mechanism results in scaled non-conformity scores Si = yi − ˆf(xi) divided by this scaling function sigma(x), and then the final prediction interval is constructed using these scores.

Tom: It’s a very specific mathematical pipeline where every piece—from the initial classification uncertainty to the final interval width—is mathematically linked through that cascade effect.

Jane: The summary really boils down to using the uncertainty from one task to make the prediction of another task more reliable by dynamically scaling the confidence measure.

The paper's improvements: Tom: Now that we’ve seen how it works, let’s shift focus to what they suggest as improvements in "CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support."

Jane: The paper suggests the main improvement is moving away from static, global uncertainty quantification toward a dynamic, context-aware scaling mechanism.

Lu: They propose integrating epistemic uncertainty from the initial classification stage directly into the prediction interval of the downstream regression stage to create that cascade effect.

Meng: I think this is where it gets practical; instead of just calculating one big uncertainty number for everything, they want a situation-specific adjustment based on how uncertain we are right now.

Lalam: They suggest using Venn-Abers predictors to get the multi-probabilistic interval in Stage one and deriving the uVA(x) score as a rigorous proxy for the stability of that clinical decision.

Tom: And then they propose employing the Continuous CASCADE approach for adaptive calibration, which uses that uVA(x) to define a continuous scaling function sigma(x).

Jane: Specifically, they suggest defining sigma(x) as one plus beta times the difference between uVA(x) and the average uncertainty across the calibration set.

Lu: They also propose scaling non-conformity scores by this function, Si = y i − ˆf(x i) divided by sigma(x i), which ensures that when uVA(x) is high, the interval expands, and when it’s low, it shrinks.

Meng: The benefit here is twofold: first, you get highly efficient intervals for confident patients because they can be up to thirty-eight point nine percent narrower than standard baselines.

Lalam: And second, when uVA(x) is high, the interval expands significantly—for example, it expands by over one hundred fifty-eight point nine percent in one comparison, which actually increases coverage from eighty-five point four percent to ninety-one point seven percent.

Tom: So they are showing that this dual effect of sharpening intervals for clear cases and expanding them for ambiguous ones is a significant feature of the CASCADE framework.

Jane: It’s about using uncertainty not just as a label but as an active control signal to guide the prediction interval construction in a way that is context-aware.

Conclusion: Tom: So we’ve gone through the whole paper, and now for the wrap-up of "CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support." Let’s summarize the big implications before we go.

Jane: Basically, this work shows how to couple classification uncertainty with regression uncertainty to create prediction intervals that are dynamically scaled based on that initial classification result.

Lu: The implication is that we can move toward more nuanced clinical decision support systems where the confidence in a dosage recommendation is directly reflected in the interval width.

Meng: For practical use, this means clinicians get tighter, more efficient predictions when they are certain about a dose change and explicit warnings when the system finds itself in an ambiguous situation.

Lalam: This framework has potential to improve AI culture by showing that uncertainty isn't just noise; it’s a signal that should be used to drive decision-making caution, making the AI more responsible.

Tom: It really moves the conversation from just building accurate models to building systems that are inherently better at handling real-world complexity.

Jane: This paper provides a solid foundation for how we can build more trustworthy sequential decision support tools in complex medical fields by integrating uncertainty directly into the prediction structure.

Lu: If you think about it, this work suggests that we can use this framework to embed uncertainty as an active control parameter rather than just reporting it passively, which is a really creative way to handle model reliability.

Meng: It’s promising because it moves us toward systems where the uncertainty is used actively to guide caution, which aligns perfectly with how complex medical decisions actually need to be made in practice.

Lalam: And for culture, this means we can develop AI that is not just accurate but also communicative about its own limitations and when it needs human oversight.

Tom: So, to wrap up our discussion on "CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support," we’ve seen a method that ties upstream uncertainty directly into downstream prediction scaling to create those adaptive intervals.

Jane: It’s a solid piece of work that offers a clear path toward more trustworthy clinical tools by making the confidence level explicit in the prediction structure.

Lu: It opens up avenues for building more sophisticated sequential decision support systems by using uncertainty as an active parameter to guide the prediction interval construction, which is a really creative way to handle model reliability.

Meng: This approach offers tangible benefits in terms of efficiency and safety, showing that we can achieve better performance without sacrificing the coverage guarantees required in clinical environments.

Lalam: It’s an important step toward developing AI that is not just accurate but also communicative about its own limitations and when it needs human oversight.

University of Florida · MD Anderson Cancer Center

cs.LG, stat.ME, stat.ML

Submitted: 2026-05-19

Updated: 2026-09-30

Comments: Accepted to ICML 2026 AgenticUQ Workshop. 14 Pages, 3 Figures

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Effective medication management in Parkinson’s Disease (PD) is challenging due to heterogeneous disease progression, variable patient response, and medication side effects.

Key concepts

Conformal Prediction
A statistical method used to provide prediction intervals that guarantee a certain level of coverage (e.g., 80%). It ensures that the true outcome falls within the predicted range with a specified probability, making it reliable for clinical decision-making.
Epistemic Uncertainty
This refers to the uncertainty in a model's predictions due to a lack of knowledge or data, rather than just random noise. In CASCADE, it measures how 'confused' the initial classifier is about whether an intervention is needed, serving as a signal for how reliable the second prediction stage will be.
Venn-Abers Uncertainty Score (uVA)
This score quantifies the ambiguity of a binary classifier by measuring the length of its multi-probabilistic interval. A longer interval means the model is less certain about whether an intervention is required, making it a direct proxy for clinical decision stability.
Continuous CASCADE Scaling
A method that uses a continuous function to dynamically adjust prediction interval lengths based on the uVA score. If uncertainty (uVA) is high, the interval expands to be safer; if uncertainty is low, it shrinks to be more efficient. This ensures intervals are optimized for both accuracy and clinical caution.

Terminology

Summary

Effective medication management in Parkinson’s Disease (PD) is challenging due to heterogeneous disease progression, variable patient response, and medication side effects. CASCADE introduces a novel conformal prediction framework that propagates epistemic uncertainty from a screening classifier to adapt downstream predictions. This approach bridges the gap between discrete clinical decision-making and continuous dose forecasting by dynamically scaling prediction intervals based on upstream classification reliability.

The gist: CASCADE introduces a novel conformal prediction framework that propagates epistemic uncertainty from a screening classifier to adapt downstream predictions, achieving continuous risk adaptation by mapping VennAbers multi-probabilistic uncertainty directly to non-conformity scores.

How it works

CASCADE is designed for two-stage architectures where the first stage identifies if an intervention is needed (classification) and the second stage predicts how much change is required (regression). The core principle involves cross-task uncertainty transfer: the epistemic uncertainty of Stage 1, which identifies whether a patient requires medication changes, serves as a critical signal for the reliability of Stage 2. This propagates reliability directly into the conformal calibration process.

The framework operates in two phases:

  1. Epistemic Uncertainty Quantification (Stage 1): This phase quantifies the ambiguity of the binary classifier using Venn-Abers predictors, which provide a multi-probabilistic interval ([p0(x), p1(x)]). The Venn-Abers Uncertainty Score, denoted as uVA(x), is defined as the length of this probability interval: uVA(x) = p1(x) − p0(x). This score serves as a rigorous proxy for the stability of the clinical decision.

  2. Adaptive Conformal Calibration (Stage 2): In this phase, the regression prediction intervals are scaled based on the Stage 1 uncertainty. The framework proposes two distinct strategies:

Mondrian (Discrete Stratified CP)

This approach partitions calibration data into disjoint strata based on fixed quantiles of uVA(x). The prediction interval is then constructed using only the residuals of calibration examples in that same uncertainty stratum, ensuring validity within each stratum.

How it works (Continued)

Continuous CASCADE (Mean-Centered Scaling)

To avoid discretization artifacts and sample fragmentation inherent in discrete methods, Continuous CASCADE uses a continuous scaling mechanism based on Normalized (Locally Scaled) Conformal Prediction. It defines a scaling function σ(x) that adapts the interval length relative to the population’s average uncertainty, parameterized by a sensitivity factor β: σ(x) = 1 + β [uVA(x)/u¯VA − 1]. This results in scaled non-conformity scores (Si = yi − ˆf(xi) / σ(xi)), and the final prediction interval is constructed using this scaling factor. This ensures that when uVA(x) > u¯VA, the interval expands, and when uVA(x) < u¯VA, it shrinks.

Evaluation Metrics

The framework's performance is assessed using several metrics to quantify its efficacy in balancing coverage and efficiency:

  1. Marginal Coverage (1 − α): This measures the proportion of test samples where the true outcome lies within the predicted interval, targeting a nominal level of 80% (α = 0.2).

  2. Average Length (Efficiency): This metric tracks the mean size of the prediction interval, aiming for shorter intervals when confidence is high.

  3. Cascade Ratio (CR) (Adaptivity): Defined as the ratio of average interval length in the highest uncertainty partition to that of the lowest uncertainty partition: CR = E[len(Cˆ(x)) uVA(x) ≥ Q1−1/K] / E[len(Cˆ(x)) uVA(x) ≤ Q1/K]. A higher CR indicates superior conditional adaptivity.

Key Findings and Validation

The validation demonstrated that CASCADE produces highly efficient intervals for confident patients, achieving a 38.9% narrower interval length compared to standard conformal baselines. Furthermore, the framework exhibits significant adaptivity:

Adaptivity Signal:

For low-uncertainty patients, CASCADE tightens the interval by 38.9%. For high-uncertainty (ambiguous) patients, it expands the interval significantly (e.g., +158.9% in one comparison), simultaneously increasing coverage from 85.4% to 91.7%, and explicitly signaling that ambiguous cases warrant additional caution before adjusting medication.

Statistical Significance:

The Kolmogorov-Smirnov (KS) Test yielded a statistic of D = 0.62 (p < 10−54), confirming that CASCADE fundamentally alters the prediction structure, shifting from a static safety margin to a dynamic distribution. The Spearman Correlation between interval length and the VennAbers score (uVA) was found to be near-perfect at ρ = 0.

Improvements for AI systems

Based on the provided scientific paper, here are specific improvements that can be implemented in existing or future AI systems:

The core improvement lies in moving from static, global uncertainty quantification to a dynamic, context-aware scaling mechanism.

  1. Enhance two-stage clinical decision support systems (like those for Parkinson’s Disease medication management) by integrating epistemic uncertainty from the initial classification stage directly into the prediction interval of the downstream regression stage.

  2. Implement a Cascade Effect where a classifier's confidence (or lack thereof) dynamically scales the reliability and width of its subsequent dosage prediction intervals.

  3. Utilize Venn-Abers predictors to quantify epistemic uncertainty in binary classification tasks, yielding a multi-probabilistic interval, and derive an epistemic uncertainty score, denoted as the Venn-Abers Uncertainty Score (uVA(x) = p1(x) − p0(x)).

  4. Employ the Continuous CASCADE approach for adaptive calibration. This involves defining a continuous scaling function:

  5. σ(x) = 1 + β [uVA(x)/u¯VA − 1], where u¯VA is the mean uncertainty across the calibration set, and β is a tunable sensitivity parameter (optimized at 0.7).

  6. Scale non-conformity scores by this function: S i = y i − ˆf(x i) / σ(x i).

  7. Construct the final prediction interval as: Ĉ(xnew) = ĥf(xnew) ± Q1−α · σ(xnew), where Q1−α is the (1-α)-th empirical quantile of the normalized scores across the entire calibration set.

This improved AI system can perform the following specific functions:

  1. Predict precise medication adjustments (e.g., LEDD percentage change) with highly optimized intervals for patients where it is very certain about a dosage change, leading to significant reduction in interval length (up to 38.9% narrower than standard baselines).

  2. Automatically expand prediction intervals when the upstream classifier is highly uncertain (i.e., near the decision boundary), providing an explicit adaptive safety buffer that signals high risk for complex or ambiguous cases, thereby mitigating risks like levodopa-induced dyskinesia (LID) or over-medication.

  3. Provide a continuous, clinically interpretable signal of model confidence that directly correlates with the uncertainty of the clinical decision, allowing clinicians to distinguish between robust forecasts and statistical guesses in real-time.

  4. Optimize the trade-off between safety and efficiency by using a tunable parameter (β) that allows practitioners to explicitly calibrate how sensitive the system should be to uncertainty, ensuring that high adaptivity is maintained without violating coverage guarantees.

Abstract

Effective medication management in Parkinson's Disease (PD) is challenging due to heterogeneous disease progression, variable patient response, and medication side effects. While AI models can forecast levodopa equivalent daily dose (LEDD) as a measure of medication needs, standard uncertainty quantification often fails to communicate the reliability of these predictions, treating high and low confidence clinical decisions identically. We introduce CASCADE (Calibrated Adaptive Scaling via Conformal And Distributional Estimation), a novel conformal prediction framework that propagates epistemic uncertainty from a screening classifier to adapt downstream predictions. Unlike standard conformal methods that rely on auxiliary residual regression, we leverage epistemic uncertainty from a primary classification task (identifying whether a medication change is needed) to dynamically scale the prediction intervals of a secondary regression task (predicting how much change). By mapping Venn-Abers multi-probabilistic uncertainty directly to non-conformity scores, our framework achieves continuous risk adaptation. We demonstrate that this cascade effect produces highly efficient intervals for confident patients (38.9% narrower than standard conformal baselines) while automatically expanding intervals to ensure robust coverage for uncertain cases, bridging the gap between discrete clinical decision-making and continuous dose forecasting in PD.

Sources

Related papers