Diagnostic Improvement through Qualitative Sensitivity Analysis and Aggregation

Keith L. DowningAAAI 1987

Introduction

Automated diagnosis has historically been divided into two distinct approaches: experiential "shallow" systems and first-principles "deep" reasoning. Shallow systems, often built as sets of if-then rules from human experts, are computationally efficient but struggle with novel or multiple-fault scenarios and offer limited explanatory power. Conversely, deep reasoning uses physics-based or physiological models to trace faults from first principles. While robust and highly explainable, these models are often too slow for real-time diagnostic applications.

Keith Downing's research introduces QUALSA (Qualitative Sensitivity Analysis), a method designed to bridge this gap. By performing "knowledge compilation," QUALSA automatically derives streamlined symptom-fault associations from complex causal models. This process allows a system to retain the explanatory depth of a first-principles model while achieving the speed of an expert system.

The QUALSA Framework

The core of QUALSA is the transformation of quantitative models into qualitative ones. Instead of tracking precise numerical values, the system utilizes "confluences"—qualitative differential equations that monitor whether a value is increasing (+), decreasing (-), or remaining stable (0). The diagnostic logic is built through a multi-step process:

  1. Fault Planting: The system manually alters a system property (e.g., increasing vascular resistance) within the model.
  2. Constraint Satisfaction: It uses an algorithm to determine how dependent variables (symptoms like blood pressure or flow) react to that specific change.
  3. Sensitivity Mapping: The system calculates "Qualitative Sensitivities," identifying whether a property affects a variable directly, inversely, or not at all.

Case Study: The Cardiovascular Model

To demonstrate the efficacy of QUALSA, the author applied it to a quantitative cardiovascular model. This model simulates the human circulatory system as a mechanical-electrical circuit. In this abstraction, properties like heart contractility and vascular resistance are mapped to electrical components, while physiological variables like cardiac output and arterial pressure are treated as current and voltage.

Figure 1: Circuit representation of the cardiovascular model, where $P_0$ is maximal heart pressure, $G$ is heart contractility, $R$ is total circulatory resistance, $Q$ is cardiac output, and $P$ is arterial pressure.

Structural Abstraction and Parameter Aggregation

One of the most powerful features of QUALSA is its ability to perform parameter aggregation. During the sensitivity mapping phase, the system may discover that two distinct properties produce identical qualitative effects on all observed symptoms. In such cases, these properties are redundant from a diagnostic perspective.

In the cardiovascular model, QUALSA identified that "Heart Contractility" ($G$) and "Maximal Heart Pressure" ($P_0$) affected arterial pressure ($P$) and cardiac output ($Q$) in exactly the same way. As shown in the sensitivity table below, both $G$ and $P_0$ have identical qualitative signatures.

Figure 3: Qualitative sensitivity table showing how changes in properties ($G, P_0, R$) affect variables ($P, Q$). The identical columns for $G$ and $P_0$ indicate they can be aggregated.

By merging these into a single abstract property—labeled "Heart Strength" ($H$)—the system simplifies the diagnostic search space without losing accuracy. This "structural abstraction" allows the system to effectively discover higher-level physiological concepts from lower-level mechanical descriptions.

Results and Implications

The resulting "Aggregated Fault Table" provides a direct, near-instant mapping from symptoms to faults. For example, an increase in arterial pressure combined with a decrease in cardiac output can be immediately diagnosed as an increase in total circulatory resistance ($R$).

Furthermore, QUALSA demonstrates robustness in handling multiple faults. By intersecting the possible interpretations of individual faults, the system can resolve ambiguities and narrow down the global state of the system even when qualitative data is sparse. While the method relies on the accuracy of the initial causal model and can face non-determinism in complex qualitative arithmetic, it represents a significant step toward automated systems that can learn and simplify their own internal reasoning for better diagnostic performance.

Summary updated

BibTeX
@inproceedings{downing1987diagnostic,
  author = {Keith L. Downing},
  title = {{Diagnostic Improvement through Qualitative Sensitivity Analysis and Aggregation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure