In health outcome research, functioning is measured from different perspectives. In clinical research, functioning is assessed to describe the limitations and restrictions of patients before and after an intervention. In the field of quality-of-life research, functioning is assessed from the patients' perspective describing how patients feel about those limitations and restrictions. In economic evaluations, the personal value that the patients attribute to such limitations and restrictions is analyzed [1].

However, functioning is not only an outcome on end results of health services in relation to chronic health conditions. At the individual level, functioning represents the starting point from which to plan interventions [2]. At the institutional and social levels, functioning provides the basis for predicting need of care [3,4], length of treatment or hospitalization [5-10] and planning the distribution of resources.

The International Classification of Functioning, Disability and Health (ICF; [11]) is the framework developed by the World Health Organization (WHO) to describe functioning and disability at both the individual and population levels. The ICF represents the universal language of functioning to be used not only in outcome assessment, but also for planning health interventions and resources. Moreover, the ICF is intended to be used in multiple sectors that include, beside health, education, insurance, labour, health and disability policy and statistics [12].

The development of the ICF was guided by a bio-psycho-social or integrative model of functioning and disability. Based on this model, functioning, with its components Body Functions and Structures and Activities and Participation, is seen in relation to the health condition under consideration, as well as personal and environmental factors (Fig. ​(Fig.11)[11]. Functioning denotes the positive aspects, and disability denotes the negative aspects of the interaction between an individual with a health condition and the contextual factors (environment and personal factors) of that individual. Thus, disability is an umbrella term for impairments, limitations in activities and restrictions in participation. This distinction can help when reading the medical literature. Disability is usually the preferred term. However, functioning is implicitly addressed when disability is studied and vice versa [13].

In 2001, 191 member states of the World Health Organization agreed to adopt the ICF as a basis for the scientific standardization of data on health and disability worldwide [14]. Since then, there is growing interest and even enthusiasm in the application of this classification in clinical practice and research [15,16]. However, it is now well understood that there are a number of challenges that need to be addressed to fully implement the ICF in everyday practice.

The ICF Core Sets represent one approach to operationalize the ICF for clinical practice and research. In the form of short, generally-agreed-on lists of ICF categories relevant for specific health conditions, the ICF Core Sets address a number of challenges that include the feasibility of the ICF and the links between functioning and specific health conditions, i.e. links between the ICF and the International Classification of Diseases (ICD-10) [16].

However, while the condition and setting- specific ICF Core Sets are useful when describing and classifying functioning for patients with specific health problems or in specific care settings, a Generic ICF Core Set is needed to describe and compare across health conditions. In other words a Generic ICF Core Set serves as a common currency and is a true operationalization of the ICF model as it applies the principle of etiologically neutrality. A Generic ICF Core Set will be developed in an iterative process involving a number of criteria and methodological approaches. One of these approaches will be the examination of the explanatory power of determined ICF categories in relation to external standards across the 12 chronic conditions for which condition-specific ICF Core Sets have already been developed.

Different approaches can be used to capture broad concepts, such as functioning and health. In the literature, it is traditionally distinguished between single questions and questionnaires capturing a number of domains [17].

The use of a single question or item representing a broad construct is an important consideration when deciding which external standard to include as independent variable in regression models. In contrast to a health profile or an index, which consist of several items, a single item provides a very broad [17] but straight-forward measure about the concept under consideration. The Short-Form 36 [SF-36; [18]] is the most widely-used health-status measure. Item one of this instrument is arguably one of the most widely-investigated single items referring to health in general [19]. Moreover, it has high face validity. Therefore, item one of the SF-36 questionnaire has been used for this investigation.

The general objective of the present study is to propose a number of candidate ICF categories for a Generic ICF Core Set.

Information on diagnoses and health status as measured by the SF-36 are shown in Tables ​Tables11 and ​and2,2, respectively.

Patients with 12 different health conditions were included in the study. The largest patient group suffered from low back pain (19.2% of the patients). Osteoporosis is the least frequent diagnosis (3.4%).

About one third of the subjects had only one disease. The number of diseases suffered by the patients ranged between 1 and 13 diseases. 76.4 % of the patients reported 1 to 3 different diseases. The average number of diseases was 2.2.

In the SF-36, the lowest health status was reported for the scale Role Physical and Vitality. The highest health status was in the scales Role Emotional and Social Functioning.

In the summary scales, patients reported stronger limitations in the PHI score than in the MHI score.

28 ICF categories were considered for the initial regression model after having analyzed which categories represented a problem for at least 10% of the patients, their correlation to the item 1 of the SF-36 and the requirement of independence.

In addition to the initial model, 13 different regression models were performed by substituting the ICF categories included in the initial model with ICF categories with which they were highly correlated. Table ​Table33 and ​and44 show the results of the 14 different regression models applied. The first line of Tables ​Tables33 and ​and44 shows which ICF categories of the initial model were substituted by which other ICF categories.

Table ​Table33 displays the R² or the amount of variance accounted for by the single independent variables and the total variance in the different models. Table ​Table44 displays the parameter estimates, the estimated heteroscedasticity-consistent standard errors and the corresponding p-values for each of the dependent variables of the different 14 models.

The initial model includes eight independent variables, from which three are the fixed control variables (age, gender and number of diseases), and the presence vs. absence of the diagnosis breast cancer and osteoporosis as the main diagnosis. The ICF categories included in this first model are b130 energy and drive functions, b280 sensation of pain, d640 doing housework.

B280 sensation of pain did not correlated above 0.5 with any other variable. Therefore, this category was not substituted for any other category in the regression model conducted after the initial model.

Models II to IV were based on the variables that correlated above 0.5 with b130 energy and drive functions (see Fig. ​Fig.2).2). B130 energy and drive functions was substituted for b140 attention functions in model II, for b152 emotional functions in model III and for b144 memory functions in model IV, respectively. In models II and IV, the ICF category b164 higher-level cognitive functions was also excluded from the analyses since this ICF category correlated above 0.5 with b140 attention functions and b144 memory functions.

The ICF categories b140 attention functions and b144 memory functions did not significantly explain any variance in item 1 of the SF-36 and did not, therefore, remain in their respective models. However, other variables, such as the presence vs. absence of depressive disorder and chronic widespread pain, significantly explained 0.9 and 0.4 of the variance of the item 1 of the SF-36 and, therefore, remained in the respective models.

The ICF category b152 emotional functions significantly explained 1.2% of the variance in the item 1 of the SF-36 and remained in the model III. Model I and III included similar ICF categories. It is, however, interesting to note that d640 doing housework explained more variance in model III, in which the ICF category b152 emotional functions was included instead of b130 energy and drive functions.

Model V substituted the ICF category d640 doing housework for d450 walking, which correlated above 0.5 (see Fig. ​Fig.3).3). Since d450 walking did not correlate to d920 recreation and leisure (as d640 doing housework does), d920 recreation and leisure was also included in the model. Otherwise, the ICF category b730 muscle functions was not included in the model because this ICF category correlated above 0.5 with d450 walking.

D450 waking remained in model V, significantly explaining 2% of the variance of item 1 of the SF-36.

Models VI to X are based on model V and in the inter-correlations presented in Fig. ​Fig.3.3. In models VI and IX, d450 walking was substituted for d475 driving and in models VII and X, for d470 using transportation, respectively (see Fig. ​Fig.3).3). In model VI and VII, the ICF category d920 recreation and leisure was included. In models VIII, IX and X the ICF category d850 remunerative employment was included.

The ICF categories d475 driving and d470 using transportation did not remain in the corresponding models VI, VII, IX and X. The ICF category d920 remained in model VI, significantly explaining 1.4 of the variance of the dependent variable. D850 remunerative employment remained in models IX and X explaining significantly 1.1 and 1% of the variance of the dependent variable. Model VIII was almost identical to model V.

Model XI substituted d640 doing housework for d540 dressing and 920 recreation and leisure. Model XII substituted d640 doing housework for d620 acquisition of goods and services. Models XIII substituted d640 doing housework for d630 preparing meals. Model XIV substituted d640 doing housework for d660 assisting others. The ICF categories d620 acquisition of goods and services and d660 assisting others remained in their respective models, significantly explaining 4.7 and 0.6 % of the variance in the item 1 of the SF-36, respectively. It is important to mention that the R² of b130 energy and drive functions decreased to 0.017 when d620 acquisition of goods and services was included in the model.

The variance of the item 1 in the SF-36, the performed models account for, ranged from 22.27% in model IX to 24.0% in model IV. The control variables accounted for 3% to 3.5% of the variance in the dependent variable. However, only the variable number of diseases revealed a significant result (p < .0001) in all the models. Age and gender were not significant in any of the models.

The ICF category that explained the highest amount of variance in all the models (from 10.7% in model I to 11.8% in model V) was b280 sensation of pain.

The variables presence vs. absence of breast cancer and osteoporosis, respectively, were consistently present in all the models, explaining around 1% of the variance in the dependent variable. In all the models these two variables had positive parameter estimates, i.e., patients in these diagnosis groups with breast cancer and osteoporosis reported better general health than patients with other diseases as main diagnoses.

With the exception of d235 vestibular functions, d620 acquisition of goods and services, d660 assisting others, and the two categories in the component environmental factors, the suggested ICF categories were addressed in at least 4 of the 6 most widely used generic health-status measures [29].

With the exception of d235 vestibular functions, which was listed in none, and d660 assisting others, which was listed in only 6 of the 12 condition-specific Comprehensive ICF Core Sets that have been developed so far, all the other candidate ICF categories were listed in at least 10 condition-specific Comprehensive ICF Core Sets [15]. Since the Comprehensive ICF Core Sets were developed to guide multidisciplinary assessments [30] in clinical settings like rehabilitation, they are relatively large. Thus, Brief ICF Core Sets representing a selection of ICF categories of the respective Comprehensive ICF Core Sets and addressing only those aspects of functioning that are essential and can be recorded in clinical studies were also developed. Within this context it is important to mention that b130 energy and drive functions, b152 emotional functions, b280 sensation of pain, b730 muscle power functions, d450 walking, d640 doing housework, and e580 health services, systems and policies are included in at least 6 of the 12 Brief ICF Core Sets developed up to now.

In addition, the 13 candidate ICF categories have high face validity, if one considers that a Generic ICF Core Set will have to include ICF categories which represent problems affecting most patients irrespective of their specific health conditions. As all others ICF Core Sets developed so far [30], a Generic ICF Core Set should include as few categories as possible to be practical, but as many as necessary to be sufficiently comprehensive to describe the patients' typical spectrum of problems in functioning across conditions.

The 13 candidate ICF categories already show a considerable overlap with the 12-Item version of the World Health Organization Disability Assessment Schedule II (WHODASII) (see Table ​Table5).5). [31,32], which is an instrument with from the same conceptual basis as the ICF. Therefore, it is expected that the results from this and future work on the generic ICF Core Set will contribute to the further development of the WHODAS II, as well as to the development of ICF-based assessment instruments.

Instead of regression modelling, other methods, such as simply rank the strength of the association after using parametric or non-parametric correlation statistics (Sperman or Person Coefficients) could be considered when selecting ICF categories in relation to a broad concept, like health. However, regression modelling enables a better understanding of the relative contribution of individual ICF categories in relation to others [33].

In addition, a transparent selection process which acknowledges that within a set of highly-related variables not only the finally-selected variable, but also other variables would have similarly contributed to a model, is necessary to thoroughly study the relationship among possible candidate ICF categories. We developed a selection strategy based on the ICF structure and examined a number of alternative models, considering best possible alternatives from our selection strategy. This process does not lead to a final answer about which ICF category to consider or not, but at least leads to a selection of candidates which needs further consideration and comparison with the results of other selection strategies in developing a Generic ICF Core Set.

The importance of not relying on a purely statistical selection was demonstrated by the example b152 emotional functions. If one had simply followed the statistical modelling, b152 emotional functions would not have been selected in a final model and would not have been considered as a candidate ICF category for the generic ICF Core Set. This would be counterintuitive for most health professionals, since emotional functions in the personal experience of health professionals and patients are relevant domains that always must be taken into account to describe functioning, disability and health. According to Cieza and Stucki (2004; 29) all 6 most widely used generic health-status measures address emotional functions.

As can be seen in Fig. ​Fig.2,2, b130 energy and drive functions are highly related to b152 emotional functions and vice versa. In a model including b152 emotional functions instead of b130 energy and drive functions, the total variance explained is similar, and all the variables entered are the same. A similar example can be seen with d450 walking which, when tracked back in the selection process, was highly associated with d640 doing housework. A model including d450 walking (model V) instead of d640 doing housework (model I) provides virtually a similar answer with regard to the variance explained and the other variables included. D620 acquisition of goods and services also appears to be a suitable substitute for d640 doing housework. However, category b130 energy and drive functions lose some importance when d620 is in the model.

It is also interesting to note that d920 recreation and leisure significantly explains some of the variance in general health when neither d640 doing housework nor d450 walking is in the same model. Taking into account that all the different models explain a similar amount of variance in the end, this indicates that all these variables explain the same variance in general health as measured by item one in the SF-36. The same can be said of d850 remunerative employment. This variable remains in the model when neither d640 doing housework nor d450 walking is in it.

Pain is the most relevant independent variable in all the models. Its importance remains the same, regardless of which variables are additionally included in the model. This result is not surprising, since pain is a leading symptom and one of the key outcomes in many different chronic conditions [34-43].

Body structures do not play a significant role in any of the models when explaining general health. This is probably because the organ systems involved in the health conditions affecting the patients in this study are very diverse. The fact that environmental factors are present in only two models is, however, remarkable. Environmental factors are defined as the physical, social and attitudinal environment in which people live and conduct their lives [11]. They represent factors that have an important influence on patients' health, and one would have expected more of them in the different models. This result should, therefore, be studied in similar studies and reflects the importance of an iterative procedure and the involvement of different criteria and methodological approaches for the development of a final Generic ICF Core Set.

Concerning the control variables, only the number of comorbidities has a significant influence on item 1 of the SF-36 in these analyses. This result is consistent through all the models. Individuals with fewer diseases feel healthier.

This study also presents several limitations that require special annotation.

It has to be recognized that the significance threshold of <0.01 that was needed to deal with the vast number of variables and that was set in the second step of the selection process is low. This might lead to the exclusion of potentially valuable ICF categories from the outset. This limitation is also in line with other limitations of the study, i.e., that our data were based in the ICF categories included in the ICF Checklist. As we now know from the development of the Comprehensive ICF Core Sets, a number of other categories not included in the ICF checklist, such as b455 Exercise tolerance functions, are relevant for 10 of the 12 health conditions considered. Regarding the Brief ICF Core Sets, the categories b455 Exercise tolerance functions, d240 Handling stress and other psychological demands and d230 Carrying out daily routine are included in seven, seven and five of the Brief ICF Core Sets, respectively, but are not included in the ICF checklist. Whether such categories will be included in the definitive Generic ICF Core Set will be investigated in the future.

A major limitation of our study is the relatively low variance explained by the ICF categories. There are a number of explanations to be considered which are relevant in the process of developing a Generic ICF Core Set. First of all, our selection was based only on the categories included in the ICF Checklist. Second, personal factors are still lacking within the scope of the ICF. However, personal factors refer to variables, such as fitness, lifestyle, social background and coping styles that definitely influence and determine general health. It has to be taken into account that, in this analysis, only the variables sex, age and number of concomitant diseases were included. Future studies should also include further personal factors as relevant independent variables. Third, no interaction terms were included in the models since the purpose of the study was simply to propose a method to select ICF categories and to identify candidate ICF categories for a Generic ICF Core Set. However, it is possible that interaction terms including, for example, b130 energy and drive functions and b280 sensation of pain, contribute to general health as measured by the item 1 of the SF-36. The interaction terms should be included in future studies. Four, and probably most important, in our patient sample there was not much variance in most of the categories, i.e., in many categories most people had no limitations. According to the guidelines that establish what kind of patients have access to rehabilitation after discharge from an acute hospital in Germany, all the patients have to be able to eat and wash themselves without external support, as well as be able to move independently on the ward. Therefore, one can assume that the limitations in functioning suffered by the patients included in this study do not represent the whole spectrum of severity of limitation in functioning of patients suffering from chronic conditions.

This limitation points out an additional limitation of the study – the generalization of the results. Only patients treated in rehabilitation centers in Germany were included in this study. Thus, patients treated in acute hospitals, day clinics and outpatients and inpatients in countries other than Germany are not represented. This again emphasizes the importance of performing similar analyses with patients in different countries who are being treated in different settings. This will be possible with the data collected in the international validation study of the ICF Core Sets involving over 50 countries and 270 study centres that is being performing until the end of 2006.

It has also to be taken into account that no further operationalisation of the qualifier scale besides the broad ranges of percentages provided by WHO were used in this study. The ICF checklist in its current form contains a more detail description of the qualifiers. For example, in the component body functions 1 (mild impairment) is defined as "a problem that is present less than 25% of the time, with an intensity a person can tolerate and which happens rarely over the last 30 days". The descriptions of the ICF qualifiers were not available at the time when the data collection of the study presented here was planed and carried out. Future studies should include the actual descriptions of the qualifier scale, since they may improve the reliability and validity of the data. Within this context it is important to mention that reliability studies are being currently performed at the ICF Research Branch at the Ludwig-Maximilian University to study the psychometric properties of the qualifier scale.

The SF-36 summary scores of the persons participating in this study averaged approximately 13 and 7 points less than the German normal population on physical and mental health, respectively. There are no substantial differences when compared to the German reference population with chronic conditions [44].

The fact that patients reported greater limitations in the PHI score than in the MHI score of the SF-36 is not surprising, as only 65 of the 1039 patients in our sample had a health condition that is traditionally considered as a mental-health related condition, namely depression. It is also important to mention that the German reference population with chronic conditions does not include mental-health related conditions.

In our study, we controlled for the presence/absence of each single health condition, including depression, in all performed regression models. However, the disproportionate higher number of patients with physical-health conditions makes it difficult to generalize our results to patients with mental-health conditions.

It is still an open question whether to consider the ICF qualifiers "8 – not specified" and "9 – not applicable" missing values represents the best strategy to cope with these response alternatives. The qualifier "8 – not specified" is used when the available information does not suffice to quantify the severity of the problem and "9 – not applicable" when a determine category is not applicable to a patient. These qualifiers are very useful from a clinical point of view but they represent a 'barrier' which is difficult to overcome with parametric statistical methods from a research or statistical point of view. Therefore, alternatives to how the information currently recorded with the qualifiers 8 and 9 could be recorded without using response alternatives that are part of the qualifiers' scale that ranges from 0 to 4 should be considered.