There are two main categories of CDSSs: those oriented to assessment and those oriented to proposal. The objectives of the assessment-oriented ones are: assessment of a patient’s past, current and future status (this includes prognosis, i.e., the likely outcome of an illness); risk quantification; and classification of patients according to their functional diversity [8]. The objectives of proposal-oriented ones are: risk prevention; and definition of therapeutic goals.

We designed and developed a CDSS oriented to assessment and, more specifically, to prognosis, in a clinical-system environment which requires reasoning under uncertainty, with standard indicators (e.g., Ingestion functions—b510) used in the representation of cases. The way of giving advice can be a push action, with warnings or alarms, or a pull action, giving support when explicitly requested. In the intra-hospital stage, support is pulled during all the rehabilitation process and pushed in the initial assessment and at the end of each rehabilitation activity. In the extra-hospital stage, advice is pulled in the prognosis support and pushed at the end of the PIE.

In the style of communication and decision-making process of CDSSs applied to rehabilitation, both a consulting model and a critiquing model are used. In our case, the consulting model is used; the system provides an assessment of health future status and risks; and the final decision is taken by the health professional. With respect to the CBR used for the decision-making process, the system proposed only includes the retrieve step of the typical retrieve, reuse, revise and retain phases shown in [9].

Regarding the use of information models in rehabilitation CDSSs, vMR is still in process of improvement and evolution by VMR Project Team [10], and there are several CDSSs that use it [11]. Moreover, VMR is designed to reduce development costs and time responses in CDSSs. As a consequence, although it is not widely implemented in hospitals and there are not any tools available to facilitate its implementation (unlike the information model EN/ISO 13606 [12]), it is probably the most appropriate information model for CDSSs today.

Several classifications and terminologies are used in CDSSs. In rehabilitation, ICF is used for encoding patients’ health status and professionals’ recommendations. Recommendations are activities and changes in environmental factors suggested by professionals to improve the quality of life of the patient. ICD-10 is used for representing diseases and SNOMED CT for other attributes; in particular, SNOMED CT has the potential of getting ICF-related terms mainstreamed in clinical systems, providing more context when necessary and including semantic relationships.

In rehabilitation processes, several questionnaires are used, which evaluate functioning, disability and health. In order to solve interoperability problems among questionnaires, items can be encoded to ICF concepts following the standardization methodology proposed by Cieza et al. (2005) [13]. Difficulties in mapping clinical questionnaires to standard terminologies and ontologies in the rehabilitation domain are summarized by Ceccaroni and Subirats (2012) [14]; e.g., data from questionnaires have usually a finer granularity than ICF core set categories. That is the reason why, to summarize the health status using ICF core sets from data obtained via questionnaires administered in the PIE (see examples of questionnaires in Section 3.2), the value of some (more general) nodes of the ICF taxonomy needs to be computed from the (more specific) node’s children. This process is therefore performed in a bottom-up fashion and only the nodes with values are considered in the computation.

Several aggregating functions can be used depending on the attribute and the domain. In the physical rehabilitation domain, the average function is usually used. Another example is the pessimistic function, where if parents have an empty value, the parents’ value is computed from the worst value of their children (ICF scale ranges from 4 = complete problem to 0 = no problem).

The CDSS, whose architecture is shown in Figure 1, uses standard information models, classifications and terminologies. Data from questionnaires administered in the PIE are stored in an electronic medical record (EMR) and a vMR reader/writer converts the EMR into vMR format. The health status obtained from vMR is represented in the knowledge base in the OWL ontology language [14]. A Java library converts original data into standardized data. A selection of instances of the knowledge base is stored in the case base, which, together with the ICF and ICD-10, is used to find similar patients via an open source CBR framework: jColibri [15]. Patients’ health status information is summarized by another Java library which uses the ICF and is accessed through an Internet-based interface.

The EN 12381:2005 “European norm of health informatics—Time standards for healthcare specific problems”, described by Ceusters et al. (1998) [16], is used to provide ontological and logical basis for describing temporal issues in healthcare. There are two types of patients’ data: static and dynamic. Static variables are absolute temporal expressions as they do not change over time, for example date of birth and date of diagnosis. In the domain considered for this study, patient’s health status and professionals’ recommendations are dynamic attributes collected once a year during patients’ rehabilitation.

Two representations of cases are implemented: a point-based representation, in which a case is the state of a patient in a given year; and a series-based one, in which a case includes several states. The temporal consistency of the case base must be ensured before the addition of a new case: if there is more than one state per year, the one with more available data is chosen; and alerts are activated when introducing temporally inconsistent states to the knowledge base.

Patient similarity is actionable by the CDSS because the most similar patient’s evolution is used for prognosis. The selection of the most similar case can be supervised by the professional. In that case, most relevant categories of k patients are shown to the professional. By default, relevant categories are ICF core set categories [17]. ICF core sets are subsets of ICF that can be formed according to functioning, pathology or rehabilitation process [18]. Core sets are useful because, in daily practice, clinicians and other professionals need only a fraction of the categories found in ICF.

To compute cases’ similarity, the k-nearest neighbor (k-NN) algorithm is used. Similarity between patients can be calculated through (2), in which X and Y are two patients, Sim(X_i,Y_i) is the similarity on the i^th attribute with weight w_i. ICF core sets and ICF taxonomy are used for weight assignment. The weight given to level-4 and level-3 core set categories is 2 and 1, respectively. Therefore core set categories which are more specific in the ICF taxonomy (level 4 core set categories) carry more weight in the similarity calculation.

Diseases terms are encoded to ICD-10 following the Centers for Medicare and Medicaid Services (CMS) and the National Center for Health and Statistics (NCHS) (2012) [19] methodology. Problems have been found in the use of ICD-10 for similarity calculation between diseases and have been solved as follows:

The similarity function between diseases used in the physical rehabilitation domain is [14]:

where i₁ and i₂ are the diseases, CN is the set of all concepts in the current knowledge base, super(c; C) is the subset of concepts in C which are super concepts of c, and t(i) is the set of individual i concepts.

Prognosis of patient’s functioning status, based on evidence or previous cases, help health professionals to analyze patients’ case and improve the efficacy of their recommendations. There have already been some initiatives for extracting knowledge patterns about the evolution over time of patients who suffer from spinal cord injury (SCI) [20]. In the following physical rehabilitation scenario, we describe a real life setting where the CDSS is used for research purposes.

Chronic patients’ data in the extra-hospital phase used in the scenario are taken from a bio-psycho-social study which analyzed the evolution of patients with neurological disability [21]. This study contains data of 661 patients who suffer from SCI over a span of 4 years. In total there are 5995 series-based cases and 2943 point-based cases obtained from PIE measurements in the period 2007–2010.

As an example, Neptune is the (anonymized) name of a 51-year-old man from Barcelona who suffers from spastic paraplegia since he had a traumatic injury in 1990. In his last PIE in 2007 (2008 data are not available), his doctor recommended that he seek an employment. After completing his 2009 PIE, his doctor analyzed his health status and recommended for him a change in his activities or environmental factors. Figure 3 shows Neptune’s summary and prognosis during his PIE in 2009. The added value of using a CDSS in this scenario, apart from predicting a patient future status, is that the prognosis motivates Neptune to find an employment, as he can see the impact of this factor on his functioning. This prediction helps doctors to analyze his case and prescribe recommendations for changes in activities, participation and environmental factors.

The proposed CDSS predicts Neptune’s functioning status based on previous cases. His summary is composed of the cause of functioning limitation, his recommendations, and the four chapters of ICF: body functions, activities and participation, environmental factors and body structures. In Figure 3, indicators of the ICF spinal cord injury (SCI) core set for which data is available appear in ICF sections. Risk attributes which have a severe or grave level, or decrease by 2 levels or more, are outlined.

Classes obtained from the EMR of the hospital are mapped to vMR classes, and standard classifications and terminologies, as shown in Table 1. In the physical rehabilitation scenario, the PIE assesses:

Due to the fact that the recommendations provided by professionals are not among available data, this study assumes that changes in part of the environment and activities arise as result of health professional’s prescriptions. In the extra-hospital phase, recommendations are activities (e.g., remunerative employment) and changes in the patients’ environment (e.g., social support services) to improve their quality of life. One criterion for selecting recommendations is to choose at least one item from each type of questions of the socio-demographic questionnaire (SDQ), namely: changes in coexistence, dependence care, housing, mobility, work activity, educational level, activities, economic benefits and health services.

To standardize to ICF and summarize patient’s health status shown in Figure 4, a Java library was developed. Categories appearing in the SCI long-term context core set are marked in bold, computed values from lower level categories are underlined. Values which are not underlined are obtained from PIE questionnaires. In this example, the result of computing average and pessimistic functions for temperament and personality functions coincide. In addition, maintaining a body position has the same value as its child (maintaining a sitting position) because there is only one child with value.

Some examples of how problematic diseases (see Section 2.4) are encoded and used in similarity calculation are:

In this way it is possible to calculate the similarity between two diseases with ambiguous representation in the ICD, such as, for example, Myopathy and Other ischemic stroke; similarity which has a value in this case of 0.5, considering related classes G71 and I63.8 respectively.

Regarding patients’ similarity calculation, a 10-NN classifier was developed using the jColibri framework.

To compare and evaluate the different case representation methods, chronic patients’ original data from the physical rehabilitation scenario are converted into the different representations. Point-based and series-based representations of Neptune are shown in Table 2 and Table 3. In the point-based representation, only the prognosis of 30 patients out of 661 who suffered from SCI could be performed because of missing data. In Table 2, the values of remunerative employment in years 2006, 2007, 2009 and 2010 are 4, 3, 1 and 1, respectively. These values, encoded according to the fine granularity are represented as D_-1D_-2S (See Table 3.) because the attribute decreases one level, then two levels and finally remains stationary. The values, encoded to the medium granularity, are represented as D_-S according to the correspondence between fine and medium symbol granularities of Figure 4 and considering that a sequence of same symbols is summarized using just one instance of these symbols; i.e., D_-D_-S is written as D_-S. Finally, in the coarse granularity, these values are encoded as DS.

The most similar patient to Neptune is dependent on the symbol granularity chosen. For example, the most similar patient to Neptune is Diana, Juno and Apollo according to the fine, medium and coarse symbol granularity, respectively. However, there are cases in which the most similar case according to the three granularities is the same. One example is the case of Venus and Minerva, which are the most similar patients according to all symbol granularities; both of them have stationary values in most of their attributes and only data from two years are available.

A weakness of Table 3 is the fact that only changes are considered and no values. This is solved in Table 4 by adding the value of the initial point. For instance, although the attribute preparing meals is stationary for Neptune, Diana, Juno and Apollo; Neptune and Juno have complete difficulty while Diana and Apollo have no difficulty. In order to include values in series-based representation, the first point level precedes the series of interval symbols. In Table 4, Vesta is the most similar person to Neptune. Ceres is an example of a person with the recommendation of general social support services, and the most similar person to her is Mars.

When computing the most similar person among people who suffered from SCI, the number of coincidences between the coarse and medium granularities is 61%, and the coincidence in the coarse and fine granularities is 35%. This low value made us decide to generally use the fine granularity in the physical rehabilitation scenario to avoid loss of information. Furthermore, the coincidence between fine-granularity series with and without the value of the initial point is 5% only. Therefore, the inclusion of the value of the initial point has a high repercussion on the similarity measure, and should be included in series-based representation.

An initial limitation of the system was the high response time in computing ICD-10 similarities and in the retrieval of the most similar case in large databases. To improve the response time, all ICD-10 similarities were computed once and then stored in a database, thus avoiding the need of loading and reasoning on the 14502 classes of the ICD-10 ontology [3] in real time. Another limitation is that the cosine similarity is an edge-based approach where diseases at the same level in the ontology have the same similarity, while their semantic distance is not the same. We will study other similarity functions proposed for biomedical ontologies [22]. Furthermore, patient’s similarity calculation is currently slow implementing the cached linear case base [15] on a typical desktop computer. In the future, an optimized organization of the case base will be studied. Finally, the optimal length of time-series to be used as cases is not studied in this experimental scenario in which time-series were only up to four points; therefore large time-series might also be a limitation of the current model if they are not dealt with adequately.