Table 3 Relevance of the six CDSS-RM elements and their related CDSS design considerations

Health data become useful when combined with human knowledge and experience

1. CDSS mimic the cognitive process of clinical decision makers

(a) Expert systems can be harmonically combined with machine learning

(b) Predictive models need to be interactive and react to new info & feedback from clinicians

Clinicians look for changes over time rather than raw measurement values

2. CDSS providing recommendations with longitudinal insight

(a) Models need to include, as predictors, trends of repeated measurements

(b) The sequencial order of clinical events should be modelled

(c) The temporal distance between clinical events need to be modelled

Data availability varies in different decision points. Data is used accordingly with varying degrees of certainty

3. Contextually realistic model performance

(a) Up-to-date, on the fly training and testing

(b) Appropriate dimensionality reduction methods

Copying wrong decisions of historical data is not a good practice

4. ‘Historical decision’ bias is taken into consideration in CDSS design

Design approaches that are built around health outcomes

Data are used according to clinical standards & protocols

5. CDSS integrating established clinical standards & protocols

Models annotate a-priori known variables, in a semi-automated feature selection approach

A significant portion of hospital data are in non-structured formant

6. CDSS utilize unstructured data to enhance feature-set with more input variables for improved performance

Natural Language Processing methods, such as text mining