Reliable diagnoses of dementia by the naive credal classifier inferred from incomplete cognitive data
Introduction
Dementia is one of the leading causes for concern in the elderly. On the personal level it reduces the quality of life. Socially, it is one of the causes of major costs in health care for the elderly population, severely demented patients requiring constant supervision and medical care [2], [24]. The most important and widely known type of dementia is Alzheimer’s disease (AD), which accounts for approximately 50% of all types of diagnosed dementia. Vascular dementia (VD) has traditionally been considered to be the second most common cause of dementia (up to 20% of all dementias, either alone, or in combination with AD). However, in recent years dementia with Lewy bodies (DLB) has been identified as the second most common single form of dementia, accounting for over 20% of all dementias [24]. DLB has previously been mistakenly diagnosed as Alzheimer’s disease and sometimes has been confused with schizophrenia. One of the major problems in the performance of clinical trials with DLB is the correct diagnosis of the disorder [24], [36], [37], [38].
There is currently no cure for dementia, although galanthamine, rivastigmine and donepezil (all acetylcholinesterase inhibitors) have now been registered in several countries for the mild symptomatic relief of AD. An extract of Ginkgo biloba has also shown some activity in the treatment of symptoms of dementia [15]. Current research is trying to provide an early diagnosis of AD, as there is some hope that these compounds may prove more effective in treating early stages of dementia, also often termed “mild cognitive impairment”, and in preventing rather than reducing symptoms [41]. A first clinical trial has been completed and shows that rivastigmine can dramatically improve cognitive functions in DLB [23].
Several problems confront research in this field. First, not all of the available systems are sufficiently sensitive to detect the early stages of dementia. Secondly, it remains to be confirmed which tests may differentiate between different types of dementia.
The present paper addresses these two problems by coupling the power of emerging classification tools and the diagnostic capabilities of a well-targeted system of cognitive tests. We propose an automated diagnostic model that deals successfully with both the sensitivity of the methods and their differential properties.
The Cognitive Drug Research (CDR) computerized assessment system has been chosen for this study. This system has been designed to provide a valid, reliable and sensitive tool to assess cognitive functions in dementia [26], [27], [33], [36], [37], [38], [42], [43]. The system is the most widely used automated system in dementia research [26] (see Section 2.1). We have used a database describing the actual health state and the past responses to the CDR system tests for about 3400 patients (Section 2.4). Data were not collected with the specific purpose of statistical analysis, so they present a substantial amount of missing values. Missing data are a fundamental problem for machine learning methods; treating them properly is essential to draw reliable conclusions.
To overcome these challenging issues we chose the classification model called naive credal classifier (NCC, Section 2.3) [46], [48]. The NCC generalizes the well-known discrete naive Bayes classifier (or NBC [5]) to imprecise probabilities [39], a well-founded generalized framework for uncertain reasoning. The NCC models both prior ignorance and ignorance about the likelihood originated by missing data, by sets of probability distributions (see Section 2.2). This makes the NCC one of the most significant steps towards reliability and realism in classification. We can find a similar modeling of incomplete samples in Ramoni and Sebastiani’s robust Bayes classifier (RoC) [29] (see Section 2.3 for a detailed comparison). When small or incomplete samples convey only scarce knowledge on a domain, the NCC maintains reliability by providing possible partially indeterminate classifications, i.e. more than one class for a given object. For example, in the present application, the NCC may map a patient to the set {AD,VD}, which means that it is not possible to discriminate between these two diseases; all the others, however, can be discarded. In other words, the NCC transforms the imprecision detected in the data into output indeterminacy, without trying to reduce it any further by doing strong assumptions.
The characteristics of the new paradigm of credal classification, in fact, enable the NCC to make automatically reliable diagnoses of dementia. This is the first application of credal classification to dementia screening. We report a detailed empirical study, in Section 3, that analyzes the predictive behavior of the NCC on the data. (The reader interested in the results from the clinical viewpoint can find a summary in the concluding section of the paper.) The diagnostic accuracy described in past work on dementia [20] is thereby improved and we obtain up to 95% correct predictions. We also show that the system is very effective in discriminating among dementias, even between the two types that are currently only hardly distinguishable, AD and DLB. The study also compares the NCC with the NBC and the RoC. In comparison with the NCC, the predictions of the NBC appear to be unreliable for a large portion of the data. The NCC compares well with the RoC, and it avoids the overly caution exhibited by the RoC in some cases.
In summary, we deal successfully with the problem of obtaining reliable conclusions, which is fundamental for the application domain under consideration and is even more critical given the incompleteness of the database.
Section snippets
The CDR system
The International Group on Dementia Drug Guidelines has issued a position paper on assessing cognitive function in future clinical trials [10]. The working group concluded that existing testing procedures (e.g. the Alzheimer’s disease assessment scale) do not properly identify all cognitive deficits characterizing AD patients, in particular attention deficits, and has recommended that automated procedures be used alongside more traditional ones to ultimately determine whether they should
Experiments
The experiments aim at evaluating empirically the NCC. This stage involves comparing the NCC with the NBC and the RoC.1
In the following, when either the NCC or the RoC produce more than one class for an instance, we say that the classification is (partially or totally) indeterminate and the classifier is imprecise or suspends the judgment. In the opposite case, we speak of determinate classifications and precise
Conclusions
Cognitive tests for dementias are becoming more and more important, as early diagnosis seems to be the basis for coping successfully with the diseases. This paper shows that coupling targeted cognitive tests such as the CDR computerized system with a reliable classifier such as the NCC, enables very accurate automated diagnoses. The results are briefly summarized below.
One thousand six hundred and seventy-three records of patients’ data entered the first experiment, set up to detect dementia.
Acknowledgements
Marco Zaffalon would like to thank Peter Walley for his encouragement to develop credal classification and for many important suggestions and enlightening discussions. Thanks also to L.M. Gambardella and C. Lepori for their kind attention and support. The authors are grateful to two anonymous referees for useful and detailed comments. This research was partially supported by the Swiss NSF grant 2100-067961.02/1, and by the Swiss CTI grant 4217.1.
References (48)
- et al.
Censoring of outcomes and regressors due to survey nonresponse: identification and estimation using weights and imputations
J. Econ.
(1998) - et al.
Efficacy of rivastigmine in dementia with Lewy bodies: results of a randomised placebo-controlled international study
Lancet
(2000) - et al.
Robust Bayes classifiers
Artif. Intell.
(2001) Exact credal treatment of missing data
J. Stat. Plann. Inference
(2002)The naive credal classifier
J. Stat. Plann. Inference
(2002)- Abellán J, Moral S. Building classification trees using the total uncertainty criterion. In: de Cooman G, Fine T,...
- et al.
Bridging study of S12024 in 53 in-patients with Alzheimer’s disease
J. Am. Geriatr. Soc.
(1997) - Bernardo JM, Smith AFM. Bayesian theory. New York: Wiley;...
- Blake CL, Merz CJ. UCI repository of machine learning databases. http://www.ics.uci.edu/∼mlearn/MLRepository.html,...
- Duda RO, Hart PE. Pattern classification and scene analysis. New York: Wiley;...
Evaluation of cycloserine in the treatment of Alzheimer’s disease
J. Geriatr. Psychiatry Neurol.
Rational belief
Behav. Brain Sci.
A placebo-controlled, double-blind, randomized trial of an extract of Ginkgo biloba for dementia
J. Am. Med. Assoc.
Cited by (59)
Detecting mild cognitive impairment and dementia in older adults using naturalistic driving data and interaction-based classification from influence score
2023, Artificial Intelligence in MedicineNonlinear desirability theory
2023, International Journal of Approximate ReasoningUsing extreme prior probabilities on the Naive Credal Classifier
2022, Knowledge-Based SystemsCitation Excerpt :Indeed, the NB algorithm has been effectively used in practice [14–16]. Also, for Imprecise Classification, NCC has been successfully employed in practical applications [17–19]. Moreover, it is convenient to remark that NCC, due to its naive assumption, is much faster than other algorithms for Imprecise Classification and requires far less computational cost.
Improving and benchmarking of algorithms for decision making with lower previsions
2019, International Journal of Approximate ReasoningCitation Excerpt :In this study, we will assume that the subject is able to model her beliefs using a lower prevision, or equivalently, through probability bounding [18,19]. Maximality [17, §3.9.1-3.9.3, pp. 160-162] and interval dominance [20, §2.3.3, pp.68-69] are well-known decision criteria induced by strict partial orders associated with lower previsions [13]. Several authors proposed algorithms for finding maximal gambles, for example, Kikuti et al. [8], Troffaes and Hable [15, p. 336] and Jansen et al. [6].
An open toolbox for the reduction, inference computation and sensitivity analysis of Credal Networks
2018, Advances in Engineering SoftwareRandom forest to differentiate dementia with Lewy bodies from Alzheimer's disease
2016, Alzheimer's and Dementia: Diagnosis, Assessment and Disease MonitoringCitation Excerpt :Random forest is one such algorithm, developed by L. Breiman, and based on the principle of decision tree learning [15]. In the field of dementia, ensemble-learning methods have mainly been studied to classify patients with AD [13], whereas very little evidence is available on the automated discrimination between DLB and AD [12] or on the combination of different diagnostic modalities in an automated classifier. This study aimed to build a random forest classifier to discriminate between DLB, AD, and controls and to quantify the importance of (combinations of) different types of diagnostic features (i.e., clinical, neuropsychological, EEG, CSF, and neuroimaging data), with a specific focus on the role of EEG.