Semi-automatic learning of simple diagnostic scores utilizing complexity measures

Abstract

Objective: Knowledge acquisition and maintenance in medical domains with a large application domain ontology is a difficult task. To reduce knowledge elicitation costs, semiautomatic learning methods can be used to support the domain specialists. They are usually not only interested in the accuracy of the learned knowledge: the understandability and interpretability of the learned models is of prime importance as well. Then, often simple models are more favorable than complex ones. Methods and Material: We propose diagnostic scores as a promising approach for the representation of simple diagnostic knowledge, and present a method for inductive learning of diagnostic scores. It can be incrementally refined by including background knowledge. We present complexity measures for determining the complexity of the learned scores. Results: We give an evaluation of the presented approach using a case base from the fielded system SONOCONSULT. We further discuss that the user can easily balance between accuracy and complexity of the learned knowledge applying the presented measures. Conclusions: We argue that semi-automatic learning methods can support the domain specialist efficiently when building (diagnostic) knowledge systems from scratch. The presented complexity measures allow for an intuitive assessment of the learned patterns.

Medidas de Complejidad Cuantitativas para Sistemas Expertos Basados en Reglas

“…Por ejemplo, Atzmüller, Hauge, Lethbridge, Menzies, Pollo y Moores, reportan métricas de madurez relacionadas con las cantidades de reglas, conceptos y atributos, así como, algunas proporciones entre las cantidades obtenidas de cada componente [3], [16], [17], [18], [21], [23].…”

A case-based approach for characterization and analysis of subgroup patterns

“…Scoring the characteristic subgroup factors relies on an adaptation of a method presented in [5]: Given a subgroup, a characteristic factor e ∈ F , and the target concept t, a 2 × 2 contingency table is constructed -similar to the technique for identifying the supporting factors. With the given target concept t and the selector e corresponding to the given factor, two binary variables T measuring the target class cases contained in the subgroup, and E identifying the cases containing the specific selector e are constructed.…”

“…Map the qps-score to a symbolic category s using a conversion table 5. Label the selector e with the obtained symbolic category…”

Gray Box Robustness Testing of Rule Systems

“…This measure can be integrated in the previously described measure Average Number of Rules (NOR) by weighting the single rules with their complexity. Several of complexity measures focusing on the complexity of scoring rules were presented in [8].…”