Abstract
In many prediction tasks such as medical diagnostics, sequential decisions are crucial to provide optimal individual treatment. Budget in real-life applications is always limited, and it can represent any limited resource such as time, money, or side effects of medications. In this contribution, we develop a POMDP-based framework to learn cost-sensitive heterogeneous cascading systems. We provide both the theoretical support for the introduced approach and the intuition behind it. We evaluate our novel method on some standard benchmarks, and we discuss how the learned models can be interpreted by human experts.
Introduction
In medical diagnostics, data acquisition is rarely uniform. Information ranges from "cheap" data, such as patient age or basic symptoms, to "expensive" data, such as genomic sequencing or invasive biopsies. These costs are not just financial; they encompass time, patient discomfort, and clinical resources. Standard machine learning models typically assume all features are available simultaneously. However, clinical practice is inherently sequential: doctors start with low-cost tests and only escalate to expensive procedures if initial results are ambiguous.
The challenge lies in automating this "adaptive strategy"—deciding exactly which features to acquire for a specific patient to reach a confident conclusion at the lowest possible cost. This paper introduces a framework for building "cascade classifiers" that decide, step-by-step, whether to make a diagnosis or "abstain" and order further testing.
Methods
The authors frame sequential data acquisition as a Partially Observable Markov Decision Process (POMDP). Specifically, it is treated as a "purely epistemic" MDP, where taking an action (ordering a test) updates the model's knowledge without altering the patient's underlying state.
Key components of the approach include:
- Deep Q-Learning (DQN): Reinforcement learning is used to learn the "policy"—the rule deciding whether to predict a label or request more data.
- Dual-Network Architecture: The system employs two deep neural networks: $f_w$ acts as the classifier, while $r_w$ handles the rejection/selection logic by determining if the "cost-to-go" for more data is justified by the potential accuracy gain.
- Interpretable Design: To ensure clinical utility, the classifier's final layer uses a masked softmax regression. This allows clinicians to see the specific weight each observed feature contributes to the final decision.
- Pretraining Strategy: Two-phase pretraining (ICCA and p-ICCA) is used to stabilize the learning process and ensure efficient exploration of feature paths.
Results and Interpretability
The model was evaluated on two UCI Machine Learning repository benchmarks: Breast Cancer Wisconsin (Prognostic) and Mammography.
Performance and Efficiency
Both ICCA and p-ICCA versions outperformed Random Forests in accuracy-per-budget metrics. On the Breast Cancer dataset, the model achieved over 96% accuracy. A key finding was the model's handling of uncertainty: for 75% of misclassified subjects, the model had explored significantly more features than average, indicating it correctly identifies "difficult" cases even when a correct diagnosis remains elusive.
Stratification and Transparency
Unlike static models, this system is dynamic, grouping patients into different "diagnostic protocols." In the Breast Cancer dataset, the model identified distinct paths (cascades) based on initial data.
As shown in the figure above, the model's decisions can be explained by studying the branching logic between specific features (e.g., features 20, 9, and 7). This transparency allows human experts to understand the "why" behind a diagnosis, making the system a viable tool for clinical decision support.
Conclusion
This framework successfully balances diagnostic accuracy with the cost of data acquisition. By modeling the process as a POMDP, the system naturally emulates clinical reasoning. Future work will explore incorporating Monte Carlo Tree Search (MCTS) to further refine cascading logic and investigating sparser models to enhance interpretability in datasets dominated by a few highly predictive features.
Summary updated