Functions, rules and models: three complementary techniques for analyzing strength data
Juan Pedro Caraça-Valente
Languages & Systems - UPM
Campus de Montegancedo s/n
28660 Boadilla del Monte, Madrid
(+) 34 91 336 73 89
Languages & Systems - UPM
Campus de Montegancedo s/n
28660 Boadilla del Monte, Madrid
(+) 34 91 336 73 94
Campus de Montegancedo s/n
28660 Boadilla del Monte, Madrid
(+) 34 91 336 74 11
Isokinetics systems are now a leading technology for assessing muscle strength and diagnosing muscle injuries. These systems are very expensive, for which reason they should be put to the best possible use. However, the computer interfaces that now come with isokinetics systems only provide a simple graphical display of the strength data, that is, do not interpretate the data. This paper presents the first phase of the I4 (Interface for Intelligent Interpretation of Isokinetic Data) project, which output two computer systems: ISODEPOR and ISOCIN. Both applications provide simple and effective interaction with the LIDO Isokinetics Machine, implementing expertise in the assessment of the data through a knowledge representation mechanism that includes functions, rules and isokinetic models. The main difference between the applications is that while ISODEPOR was built for sports physicians, ISOCIN was built for blind physicians, accounting for their specific impairments.
Intelligent Data Analysis, Knowledge-based Systems.
The assessment of muscle function has been a primary goal of medical and sports scientists for decades. The main objectives were to evaluate the effects of training and the effectiveness of rehabilitation programmes [4,5].
The release of isokinetic systems meant that a group of muscles could be safely exercised to full potential throughout the entire range of movement. Used for diagnostic purposes, these systems can calculate the strength generated by the muscle during this type of exercise at each point along the arc of movement. This is tantamount to a complete dynamic assessment of muscle strength, which can be represented graphically.
Basically, an isokinetics machine consists of a physical support on which patients perform exercises within different ranges of movement using any of their joints (knee, elbow, ankle, etc.) and at a constant speed. The machine records the strength applied throughout the exercise.
The methods for assessing muscle strength using isokinetic techniques are well established in the field of injury rehabilitation and medical monitoring especially of top-competition athletes, as they provide an objective measurement of one of the basic physical conditions that is important in a wide range of sports.
The data measured by the isokinetic dynamometer are presented to the examiner by means of a computer interface. This interface sets out given parameters, which are used to describe the muscle function tested (e.g., maximum strength peak, total effort, etc.).
The mechanical component of the isokinetic systems now on the market generally meets the demands of muscle strength assessment. However, the performance of the software built into these systems still does not make the grade as far as isokinetic test interpretation is concerned. This means that the massive data flow supplied cannot be fully exploited. And, of course, visually impaired physicians cannot use these systems at all.
The I4 project developed in conjunction with the High Perfomance Center for Spanish Athletes in Madrid (hereinafter referred to as CAR-Madrid) and the Physiotherapy School of the Spanish National Organization for the Blind (hereinafter referred to as EFONCE), aimed at performing a more comprehensive analysis of the data output by the isokinetics machine. This involves making an intelligent analysis of the strength curves output by the isokinetic tests on which the assessments are based. Additionally, the system was equipped with a user interface and aids and appliances for the blind, by means of which to analyze graphical and numerical strength data. This provides for the results to be interpreted by blind physicians in particular.
In the following sections, the paper describes both applications in general and gives a more detailed description of the intelligent analysis mechanisms. So, section 2 describes the architecture of the application with its main modules; section 3 describes the three types of knowledge structures that comprise expertise, that is, functions, rules and isokinetic models, explaining each one's purpose in data intepretation; section 4 provides some details of the I4 application and the article concludes by discussing some conclusions and future work.
Figure 1 shows the architecture of the I4 system from the viewpoint of its functionalities. The isokinetics machine includes the LIDO Multi-Joint II system, which supplies the data of the isokinetic tests run on patients. As shown in Figure 1, after the isokinetic tests have been completed by the LIDO system, the first operation performed by I4 is to decode and transform the data files output by LIDO into a more standard format and to correct any inaccurate or incomplete particulars. This is the only I4 module that depends on the LIDO isokinetics system, which means that this would be the only module that would require changes if I4 were to be adapted to another isokinetics system.
Figure 1. I4 system architecture.
After execution of the above module, the data are transformed to an easier to handle, simpler and more efficient format, which is stored in the database for later use. The exercises stored can be displayed either individually or jointly in graphical format. So, using the display system, it is possible to analyse an individual exercise, compare the exercises performed with the left and right legs, any two exercises performed at the same angular speed or an exercise with a pattern or model that is representative of a particular group.
The intelligent analysis module is the most interesting system module and will therefore be discussed in more detail in the following section. Finally, the report generation module is responsible for editing and printing reports on the exercises and tests displayed.
In order to gain a clearer understanding of the knowledge representation mechanism described in this section, we start by giving some details of the isokinetic data concerned. Each session with a patient, that is a test, is composed of a set of exercises performed at particular constant speeds, depending on the test protocol currently in use . Each exercise is characterized by the speed (60, 180 or 300º/s) and the leg used (left or right). The patient will be in a seated position and the movement is made within a 0º to 90º flexion/extension arc of the leg. The LIDO system records the strength applied by the patient every 2/100 s, and the angle. Although at constant speed, the angle and time should be equivalent, the angle is recorded because minor deviations from the constant speed are very significant. The graphic representation of these data (strength over time) looks like a sinusoidal curve, containing a lot of small peaks and other irregularities (see figure 2). The amplitude, total area and irregularities are the main parameters in the analysis of the tests.
I4 interprets the isokinetic data through an expert system, the knowledge for which was elicited from the person who is, probably, the most experienced Spanish physician in muscle strength assessment using isokinetic systems. This knowledge is represented in the system by three different structures: functions, rules and isokinetic models. The functions include procedural knowledge, the rules include deductive knowledge, the isokinetic models include structural knowledge, all of which goes to make up the ES knowledge.
The objective of this representation structure is to assess the morphology of each isokinetic curve and to eliminate some irregularities for which the patient is not responsible. We firstly describe the tasks involving functions and then examine the functions in more detail.
It is the exercises performed at 60º/s that call for a more in-depth and thorough analysis, because they supply a huge amount of information. Firstly, the strength curves are preprocessed in order to eliminate inertia peaks, that is, maximum peaks produced by machine inertia rather than by the actual strength of the patient. This is detected when the angle at which the maximum peak is produced deviates a lot from the norm. Figure 2 shows a graph with peaks and the same graph after they have been removed.
Figure 2. Exercises with inertia peaks and their elimination.
Also in this preprocessing stage, I4 detects exercise extensions and flexions that are invalid because much less effort was employed by the patient than was in others, and movements that can be considered atypical as their morphology is unlike the others. The process for detecting these latter anomalies is used in part to detect the repeated extensions and flexions that are most representative of the exercise, that is, the extensions and flexions which provide for a better muscular assessment of the patient.
The analysis of the strength curves in the strict sense involves the assessment of different characteristics of the extension/flexion curves morphology, which are themselves of interest to the specialist and are some of the inputs for patient assessment. So, the aspects evaluated are:
· Uniformity. Estimates uniformity, meaning how similar repeated extension and flexion exercises are.
· Regularity. Estimates the regularity of the exercise, that is, whether the curve has a smooth contour or a lot of peaks.
· Maximum peak time. Outputs a qualitative value of the time it takes to reach the maximum peak for both extensions and flexions. This time is estimated on the basis of the slope up to the maximum peak.
· Troughs. Indicates the existence of troughs, prolonged drops and rises of the value of the moment of exercise extensions and flexions. Figure 3 shows an example of an isokinetic exercise with troughs.
· Shape of the curve. Evaluates the shape of the exercise curve for both extensions and flexions, by means of an exercise morphology study, taking into account the effort employed at the central angles, the flattening of the curve and the angle at which each maximum peak is reached.
Figure 3 Troughs during extension.
The analysis of these morphological aspects of the curves may seem straightforward for any experienced physician. However, their automated assessment is important for an inexperienced physician and crucial for sight-impaired physicians to be able to do their job. This was what primarly motivated the project and why I4 does not simply ask the user to enter the characteristics of the curve, as other expert system would.
Figure 4 shows the interface window that displays the results for the user. They are reported to sight impaired users by voice synthesis. Indeed, the entire interface can be used by blind people according to the very high standards that have been applied in many other applications built at CETTICO (Centre of Computing and Communications Technology Transfer) for the disabled.
A similar two-phase process is performed with the exercises at 180º/s and 300º/s. In both cases, the analysis phase looks at the set of curve characteristics that are more interesting for assessing patient muscles.
The above process is performed by a set of functions that remove invalid exercise extensions and flexions, eliminate the inertia peaks and analyze the curves to assess their characteristics. These functions implicitly contain the knowledge required.
The design and implementation of these functions can be described as interactive human induction, that is, given a number of strength curves, the expert evaluated each one and assessed its characteristics (i.e. whether it had inputs, troughs, the shape of the curve, etc.). Then, tentative functions were implemented, whose inputs were strength curves and whose outputs were the same characteristics for the given curves. These functions were applied to a new set of tests, and the results obtained were shown to the expert for evaluation. This evaluation led to some changes in function implementation, and so on. This process ends when the methods provide the correct value in a high percentage of the cases (over 98%). It took 3 to 5 iterations, depending on the complexity of the interpretation of each characteristic.
Figure 4. Window showing the morphology analysis of an exercise at 60º/second.
Another important issue for functions implementation was the need to assess each case according to the characteristics of the patient, male or female, age, sports, injuries, etc. A decision table for each characteristic was elicited from the expert to solve this problem. Each decision value is stored in the respective table and entered in the functions, according to patient characteristics.
It was obvious from the very start of the design phase that functions were not enough. They represent procedural knowledge very well, especially if it involves calculations, but they are not suited for representing fine granular knowledge, like heuristic assertions "If there are many invalid exercises, repeat the test". The most straigth forward means of representing this knowledge is to use "If ... then ..." rules. Another important feature of this formalism is explicit knowledge representation, as it provides an intuitive representation that can be easily validated by the expert.
The rule expert system provides conclusions on three aspects of isokinetics analysis:
· Protocol validation. This part of the analysis has the mission of determining that the protocol has been correctly applied. This is very important since the expertise used for the later parts of the analysis is very sensitive to the way in which the tests are performed. All the exercises must have been completed sucessfully, the patient must tire to some degree, etc. Only if this validation is correct, will it be possible to further analyze the data.
· Numerical analysis of data. Every numerical feature of the curve (maximum peak, total effort, gradients of the curve, etc. is expertly analyzed and conclusions provided to the user. There is an individual analysis of each leg and a comparison between both legs.
· Morphological analysis of data. The last part of the rule-based subsystem analysis, regarding morphological aspects of the data, takes into account the output of the expert functions described in the previous section. That is, the rules cannot evaluate the morphology of the strength curves (that is what the functions do); the rule-based subsystem analyzes the morphology of the strength curve of each leg and its comparison and tries to determine any kind of dysfunctions.
Figure 5 shows the interface window for the rule-based subsystem analysis. There are three tags for each part of the analysis described. The morphological analysis has been selected, and there are icons for displaying the conclusions regarding each leg, its flexion and extension. The bottom window shows an overall analysis for the right leg. It points out some problems with the right flexors and extensors and suggests that this could be caused by some dysfunction in the knee joint.
Figure 5. Window showing results from the rule subsystem
One of the most common processes performed to evaluate patient strength is to compare the results of their test against a standard. Yet another structure had to be built to represent expertize and assist physicians with this task. This third structure, which has been named isokinetic model, is composed of a standard isokinetic curve and a set of attributes. The curve is automatically calculated given a set of tests. The user selects a group of tests, usually belonging to patients of similar ages, same sex, similar sport if any and or same injury, and the system calculates the reference curve for that group. This requires some preprocessing in order to discard bad exercises (using functions). Then the isokinetic curve is calculated as the average of all the curves in the group.
Of course, this is a poor representation of the strength of the group of patients, since some of the significant irregularities are softened or even eliminated when calculating the average curve. A set of attributes had to be added to return the information needed to compare a new patient against an isokinetic model. Some of the attributes are calculated automatically from the set of tests performed on the group in question, like standard deviation, whereas others, like how close the new curve should be to the model (in terms of standard deviation) for it to be considered as normal, are defined by the expert.
Isokinetic models play an important role in the isokinetic assessment of patients, as they allow physicians to compare each new patient with standard groups. It is possible to compare a long jump athlete with a group of elite long jumpers, compare a promising athlete with a set of models to determine which is the best suited discipline, assess strength dysfuntions in an apparently normal patients, etc.
One of the reasons for implementing the ES with three different structures is that each comprises the different types of knowledge provided by the expert more intuitively. The individual tasks needed to perform a full isokinetic analysis of the patient required different knowledge representation structures. The knowledge comprised in each structure (functions, rules or models) not only can be used separately, to provide conclusions on particular issues, they can also be employed together to provide more general conclusiones.
As we have already mentioned, these three types of knowledge are related and are complementary to each other: the rule-based subsystem needs the results supplied by running the expert functions on the curves. The output of these functions is an important input for the rule preconditions. Functions also play a major role in the creation of isokinetic models, as they provide the values of some model attributes. These attributes are intended to reflect some of the common characteristics of the strength curves that could be eliminated when calculating the average strength of the individuals of a group. The conclusions supplied by the rule-based expert system is also important for this task.
However, the relations between these structures for representing expertise are not one way. The models also input facts to the rule-based subsystem for assessing comparisons between a patient and a model. Numerical (a patient is 20% stronger than the reference group) and morphological comparisons are inputs to the rule-based subsystem, which provide a higher level conclusion than a diagnosed dysfunction or any other relevant conclusion.
Therefore, the definition of three different structures for representing expertise is important, for two main reasons:
· The isokinetic domain (like most domains) contains different sorts of knowledge, each of which is better suited for a particular knowledge representation structure.
· The cooperation between the three knowledge representation structures is able to supply higher level conclusions that would be difficult to achieve using only one knowledge representation structure.
The I4 system described in this paper was designed using object-oriented methodology . Exercises, tests and models are the most significant classes of the object model in relation to the interpretation of data.
I4 has been implemented in a visual environment according to a prototyping methodology, and its first phase is now fully operational. This methodology is especially suited for this project, as it encourages interaction between software development and medical/sports assessors. The project has produced two applications, which are very similar, except for one thing: the interface. As discussed earlier, the ISOCIN application  is designed for use by sight impaired physicians, so the interface includes complete voice synthesis of every piece of information presented to the user. This includes information on how to use the system, the options open at anytime and, of course, the isokinetic data and their interpretation. It is possible to use the system without a display. ISOCIN is currently being used by blind physicians at the EFONCE to analyse injuries and assess their evolution, adapting the physiotherapy administered and rehabilitation.
The other application, ISODEPOR, is being used at the CAR-Madrid to evaluate the muscle strength of Spanish top-competition athletes.
An extract of the rating of the ISODEPOR system by physicians at the CAR-Madrid is given below. It lists how the use of ISODEPOR has improved the work of isokinetic physicians :
· Friendly access for the medical practitioners working with these systems to the isokinetic parameters in routine use, and an improved graphical presentation of the results of isokinetic tests, making reports easier to understand and more useful.
· Full analysis of the isokinetic strength curve, from which the complex or specific strength parameters that are of use for interpreting the tests can be inferred more completely.
· Output of standardized isokinetic curves by population groups (by sports, specialities, diseases, etc.), used to detect, by means of comparison, slight deviations from group norms, which raises the diagnostic potential of the isokinetic system used.
· Intelligent analysis of the strength curves obtained from the isokinetic tests that provides useful evaluations. Extraction of higher quality information for outputting reports, which make better use of the systems features.
· Use and exploitation of the system by physicians that are not specialists in isokinetics, thanks to the help provided by the intelligent interpretation.
The I4 system will provide more knowledge of the characteristics of athletes' strength, which has implications for the development and evaluation of training and rehabilitation programmes. The deployment of the I4 interface is a major advance in isokinetic data processing, as it means that muscle strength measurement systems can be better exploited. These issues make it highly relevant in the field of top-competition sport.
The intelligent analysis module is particularly useful for specialists, as it facilitates some of their work. However, as there are very few specialists in isokinetic assessment of muscle strength data, this system will be extremely valuable as an instrument for disseminating isokinetic technology and encourage non-expert medical practitioners to enter this field.
Society today has at its disposal specialized technology in a whole range of activities. This technology outputs a huge collection of information that cannot be interpreted by a non-specialized user and whose understanding is not always a straightforward matter for a specialist. There is, therefore, a demand for expert systems that interpret this information by means of an intelligent analysis of the data output. The goal of the I4 project was to provide one such system, focused on the field of isokinetics.
For this purpose, we have developed an intelligent system, built by means of an incremental KBS development model, which has the following peculiarities:
· A modular system architecture: a) data decoding, testing and storage; b) intelligent data analysis; c) data and results display and report generation.
· An expert system for the intelligent analysis of data, which structures procedural knowledge by means of functions, declarative knowledge by means of rules and structural knowledge by means of an isokinetic model.
The model has been built into two applications, one for the blind (ISOCIN) and another for elite athletes (ISODEPOR), the results of which have been highly praised by isokinetic specialists.
Although both applications were complete at the time of writing, there are still a lot of unsolved problems in the field of computerized isokinetic data analysis. Indeed, the following points are now in the process of being defined:
· Identification of patterns in the strength curves associated with each type of muscle and injury, which would enable automatic injury diagnosis and recovery analysis;
· The possibility of putting sound to the strength curves is under study, with a view to improving ISOCIN use by blind physicians and physiotherapists.
The authors would like to thank Africa Lopez-Illescas for her expertize and collaboration. The I4 project is being sponsored by the CICYT (Spanish Comission for Science and Technology), through contract number TIC98-0248-C02-01.
 Alonso, F., Barreiro, J., Fuertes, J., Martínez, L., Montes, C., 1998. Incremental Prototyping Approach to Software Development in Knowledge Engineering, Proceedings of the 5th Int. Conf. on Information Systems Analysis and Systemics, Orlando.
 Alonso, F., Barreiro, J., Valente, JP., Montes, C., 1998. Intelligent Interpretation of Strength Data, 11th International Conf. on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, Benicassim, LNAI nº 1415, vol. I, Spain.
 García, C., Zanoletty, D., 1998. ISOCIN. Sistema Experto para la Interpretación de Isocinéticos. Technical Report. Facultad de Informática, Madrid.
 Gleeson, N.P., Mercer, T.H., 1996. The utility of isokinetic dynamometry in the assessment of human muscle function. Sports Medicine, Vol. 21 (1).
 Kannus, P., 1994. Isokinetic evaluation of muscular performance: implications for muscle testing and rehabilitation. Int. J. Sports Medicine, Vol. 15, 1.
 López-Illescas A., 1993. Estudio del balance muscular con técnicas isocinéticas. In: Proceedings of the I Curso de Avances en Rehabilitación, Seville.
Copyright 2000 ACM
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy
otherwise, to republish, to post on servers or to redistribute to lists,
requires prior specific permission and or fee.
SAC 2000 March 19-21 Como, Italy
(c) 2000 ACM 1-58113-239-5/00/003>...>$5.00