Diagnosis Systems for Dementia

Medical diagnosis systems already exist and assist physicians and psychologists worldwide. However in 2012, no diagnosis system exists for dementia. While a fast and precise diagnoses can contribute to better care for the patient and more certainty for the caregiver. This project is an exploration of suitable AI techniques for such a system. Furthermore it looks at the practical side of a diagnosis system for dementia by making use of the knowledge of an expert in the field.

The diagnosis of dementia

Dementia is a cluster of different phenomena, particularly the deterioration of memory and other cognitive functions. It is a very frequent condition that strongly increases with age. The diagnoses procedure consists a couple of steps (Fig 1) and is adjusted to the client and his/her surroundings. The psychologist collects information for the diagnosis from observations, existing documentation and stakeholders. An artificial diagnosis system can contribute by assisting the psychologist with new and processed information.

Goal

The project answers the question: Can artificial intelligence techniques be used for the automated diagnosis of dementia? The focus of this project the exploration of four suitable and prominent artificial intelligent techniques for a diagnosis system. These are:

• Logistic regression
• Nearest neighbor
• Neural networks
• Support vector machines

The techniques are selected for this exploration, although other techniques also might suit for this diagnosis problem. Furthermore it will look at the practical side of a diagnosis system for dementia, by using knowledge and experience of an expert.

Methods

Data acquisition and pre-processing

For training and testing the models, data from two different institutions was used: academic hospital UMC St. Radboud (651 clients with 192 attributes) and care institution De Zorggroep (19 clients with 87 attributes). De Zorggroep dataset contains far less subjects compared to UMC St. Radboud dataset, because every client was manual extracted from a screening rapport.

For combining the two datasets, the trivial attributes were made uniform. Additionally, the following steps were made:

• Six relatively straight-forward transformations have been performed. For example, categorisation of education.
• Twenty of the attributes were not present in one or both datasets and have been computed to make the datasets complete. For example: normalisation from date-of-birth to age.
• Removal of attributes, which were not present in both datasets or not relevant to the diagnoses task. For example: NaasteInBeeld.

The combined dataset contained: demographic data, neuropsychological test results and the diagnosis.

Neuropsychological tests

The Cambridge Cognitive Examination (CAMCOG) is a commonly used neuropsychological test and is a backbone of the Cambridge Mental Disorders of the Elderly Examination (CAMDEX). The aim of the CAMDEX is to design a single, standardised instrument for accurate clinical diagnosis of dementia and the dementia classifications. Administering the CAMCOG takes about half an hour.

The aim of the Mini-Mental State Examination (MMSE) is the detection of mental status changes, particularly in the elderly and thereby enhances patient care. The MMSE is usually tested together with the CAMCOG.

The Weka workbench

For training and testing classifiers the project used version 3.6.7 of the Weka workbench. The Weka workbench is a collection of machine learning algorithms and data pre-processing tools. The classifiers used in Weka are: Logistic, Multilayer Perceptron, NNge and SMO.

Results

The performance of the classifiers is near 80 % (Fig 2). The expert is not surprised to hear this result: the classifiers use only the test results and some personal details and not all the data that the psychologist consumes. The psychologist explains:

Various factors can influence the test results. A psychologist takes into account whether the client has a permanent or temporary handicap. Visual, auditory, phatic or motor impairment issues can make the test more difficult. There are also other circumstances, like CVA (stroke), that decrease the cognitive functions. Depression or even depressive symptoms decrease the CAMCOG score. In those cases, the low scores are not caused by dementia.

The four classifiers do not differ al lot in performance, and give most of the time the same predictions (83.33 %)(Fig 3). The performance does not improve, when using plurality voting with the four or three best classifiers. An explanation for this could be that in most cases the classifiers make the same prediction, even when they are incorrect.

Evaluating individual subjects

In order to investigate the 80 % ceiling, the classifications of the De Zorggroep were discussed with the expert. The classifiers predicted most instances correct. Only in four cases one or more classifiers predicted incorrect. Interestingly, instance 3 and 7 were classified incorrectly by all classifiers.

The expert reconstructed the diagnosis of these four instances. We note the following:

• Instance 2 -> The expert confirms no dementia diagnosis, because the cognitive impairment was caused by a CVA.
• Instance 3 -> It could be possible that the classifiers did not recognise the dementia, because it is still in an early phase.
• Instance 4 -> Subject 4 was originally diagnosed with dementia. The expert disagrees with the co-worker and diagnoses subject 4 with no dementia, but the cognitive disorder NOS (DSM IV).
• Instance 7 -> The subject has no dementia, because the cognitive impairment was caused by a CVA. Also the visual impairment (Hemianopsia) and motor impairment probably influenced the test results negatively.

Additional attributes

As discussed previously, the CAMCOG is not always conclusive. Human observations could exclude other causes of abnormal cognitive functions. In order to validate these theorems, additional attributes were added to the De Zorggroep dataset: 5 mood related, 4 impairment related and CVA. However, the new dataset was too small for training the classifiers. The results of this round were not significant (p > 0.1).

Conclusion

None of the classifiers outperforms the others. The dementia diagnosis is not a problem that solved best by one specific classifier technique. Other AI techniques could also be promising as well and should be tested in further research.

Most of the time, the four classifiers give the same predictions. In the evaluation together with the expert, the expert stated that the input missed some of the observations and data to interpret the CAMCOG scores. For example the classifiers could incorrectly predict dementia, if a CVA could cause low cognitive functions. Additional attributes are needed to break the performance ceiling.

When using a diagnosis system in practice, 80 % correct diagnoses is clearly not enough. Further research in possible techniques and a new dataset with the additional attributes would assist and improve the diagnosis of dementia.

Appendix

List of attributes

The combined dataset contained:

• Demographic data
  • Gender (M/V)
  • Age (numeric), 3 age categories (boolean) and age category (string)
  • Education level (numeric) and 2 education categories (boolean)
• Neuropsychological test results
  • CAMCOG (numeric)
    • 8 subscores, 2 section scores and CAMCOG total
    • 16 median and border scores for clients gender, age and education
  • MMSE (numeric)
• Diagnosis
  • Dementia (boolean)

Suggested additional attributes

• Motivated - Is the client motivated during the examination?
• Depressed - Has the client depressive symptoms or is the client depressed?
• Visual, auditory, phatic and motor impairment - Has the client an impairment?
• Illness awareness - Is the client aware that the he/she is ill?
• Illness insight - Has the client insight in the clients’ illness?
• CVA - Has the client had a cerebrovascular accident (stroke)?

References

• Kunneman, Y. J. (2012). Diagnosis Systems for Dementia.