Abstract - The aim of this study is to examine different cognitive health risk factors that could possibly be contributing to the risk of falling within older adults in the Republic of Ireland (ROI). Possible health risk factors associated with older adults and those with dementia were derived from The Irish Longitudinal Study on Ageing dataset and were used with a machine learning approach to predict if they can be contributed towards falls. This study involved a secondary data analysis as there was no direct contact with respondents and all data was anonymized before analyzing. Risk factors were originally taken from a previous project ‘Risk Communication in Dementia’, whereby risk factors were identified and can now be used in a data analytic approach to predict risk and minimize harm. Using health risk factors such as overall health, mental health, long-term health conditions, blackouts, fainting and joint replacements, these have been proven to contribute to the risk of falling.

Keywords— Classification, Falls, Health Risk Factors, Machine Learning, Older Adults.

I. INTRODUCTION

ALLING has become part of the natural process of ageing in older adults and unfortunately this group of adults has

an increased risk of having a fall [1]. The Health Service Executive states that 30% of adults over the age of 65 years, who live in a community setting at home, will fall at least once per year [1]. Older adults who live in a residential nursing setting have a staggering 50% risk of falling. Everyone has the potential risk of having a fall, however those over 65 years old are known as a vulnerable group of people prone to falls due to cognitive decline [1]. Additionally, the fear of falling can be linked to someone having at least one previous fall, leading to more recurrent falls [13].

F

A. Health Risk Factors As risks are becoming high profile in the world of decision making, health and social care professionals are focusing more attention on risks. Decisions are made every day by professionals regarding the safety and wellbeing of older adults including those with dementia as they communicate about the potential risks they face. Risks can be viewed as positive or negative depending on how people see them; negative risks could be falling [3], burns [4], driving, wandering, forgetting about medication or taking too much medication [5]. Fig 1 displays a small amount of risk factors related to older adults and those with dementia.

In this study, the focus is placed on the negative risk of falling in older adults. Health professionals, older adults and their families should be encouraged to communicate freely about risks and work around the idea of trying to reduce risks to minimize harm [6]. Within the decision-making process there is a well-known heuristic known as a classification tree (decision) or a fast and frugal tree [7]. This study used a range of different models including a decision tree as one of the machine learning algorithms for predicting falls to establish known factors and eliminate any inappropriate factors.

B. Machine Learning AlgorithmsComputational intelligence incorporates many different

algorithms such as Neural Networks, Fuzzy Logic Systems and many more. Machine learning can be described as a collected of data analytic methods aimed at building a predictive model using an algorithm from a dataset [14]. Humans use their own knowledge and expertise to predict different outcomes every day. Machine learning algorithms can be supervised or unsupervised. Supervised can be broken down into classification or regression, Brownlee describes the process as a teacher supervising a learning process where the learning will come to an end only when the algorithm outputs an acceptable variable [15]. Unsupervised deals only with input variables to learn from the data, there is no teacher and no correct answer.

For example, a data mining approach was used to determine risk factors linked with type 2 diabetes using a decision tree and random forest approach [10]. This model predicted that the family history of diabetes was strongly associated with the development of type 2 diabetes. The random forest approach overall had a higher percentage accuracy of 71% and the decision tree was 65% accuracy proving that in this instance the random forest model was the better approach.

Figure 1: Examples of risks associated with the ageing population

Classification of Health Risk Factors to Predict the Risk of Falling in Older Adults

A Support Vector Machine was used as an example of a nonlinear classification or regression approach to computational intelligence [9]. Previously, SVM’s were implemented to separate tumors and classify them into tumor types in the diagnosis of breast cancer. This previous research has been able to reduce the computation time without losing the accuracy of the diagnoses for the patient.

Decision trees used in machine learning are becoming a popular classification algorithm allowing results or calculations to be easily understood and therefore simple to interpret [11]. The classification via Regression method performs the classification using a regression method and for every class value a regression model is built [12].

II.METHODOLOGY

The aim of this study is to combined machine learning with health risk factors from the TILDA dataset to predict the likelihood of falling. The Irish Longitudinal Study on Ageing (TILDA) is a cohort study of 8000 ageing subjects based in the Republic of Ireland (ROI). Data used in this study were taken from wave one and two which were collected over a period of time from 2009-2013. Information for the data sample was collected by carrying out face to face interviews conducted by trained professionals after consent was given by the participants. Wave one represented a random sample of people (753) who were part of the sampling framework for the study. Wave two consisted of a pre-interview questionnaire and a face to face computer assisted personal interview with the same random sample of people as wave one apart from those who had passed away. Wave one and two consists of different aspects of the lives of older people in Ireland including their health and healthcare, pensions, housing and accommodation, mobility issues, education and their employment.

The dataset was collected in three steps, firstly a face to face interview based on sociodemographic, wealth, lifestyle, social support and health. A self-declaration questionnaire was then completed and lastly a detailed health assessment was carried out by trained professionals. If participants could not leave their home to attend a self-assessment health centre, professionals attended their homes and carried out tests regarding cognitive ability, mobility, strength, bone and vision tests.

The TILDA dataset has previously been used to examine the impact of risk factors for coronary heart disease in older Irish adults. Analysis in this study was stratified by gender, age group and socioeconomic position then weighted on the related disability and compared against another dataset in Northern Ireland [8]. However, to date these data have not been used to predict falls which is the focus of this paper.

The falls attribute was originally the number of times fallen which was adapted to include a two-class classification which consisted of falls and no falls. The dataset was split into a training and testing set using both a cross fold validation of 10 and splitting the dataset into 90% training and 10% testing to

allow the model to use the training data to predict against the testing set.

Throughout wave one and two there were qualitative and numerical attributes, the numerical attributes such as falls were recorded using “How many times have you fallen in the last year?” This result was changed to a qualitative answer such as fallen and not fallen. Anyone who had fallen once or more than once was categorized into one group.

The six variables used as input factors are as follows. First was the overall health description, participants were asked to choose from excellent, very good, good, fair or poor. Secondly, to define their emotional mental health using the same answers, excellent, very good, good, fair or poor. The third variable was a yes or no answer deriving from the question ‘Have you any long-term health issues?’. The fourth variable used was a yes or no answer which asked ‘Have you previously had a blackout or fainted?’. Participants were asked if they were afraid of falling which could relate to falling more due to a fear and lastly if the participant has had any joint replacements which used a yes and no response.

Ageing and falling are becoming a popular research topic however the research does not specifically point to one attribute being the main cause of falls, its usually more than one due to the complex nature of risk factors.

The results shown in Fig 2 present the percentage of falls across the three different age groups and the slight difference in wave one and two. You will notice from the results that the percentages are spiking the higher the age group, which could represent the initial decline of cognitive function in older adults. Overall, the 50-65 age group were 57%, the 65-74 category made up 26.5% and the 75+ group consisted of 16.5% in the dataset.

50-65 65-74 75+

17.60%19.90%

24.60%

15.10%19.30%

24.80%

Percentage of Falls across Age Groups

Wave One Wave TwoFigure 1 Bar Chart Percentages of Falls

III. RESULTS

All data used was retrieved from The Irish Longitudinal Study on Ageing combining the social, economic and health circumstances of adults over the age of 50. The TILDA dataset is a nationally represented sample consisting of 8,175 adults including 4,431 women and 3,744 men. The data comprises of the various risk factors mentioned in Table 2, the data were analyzed and balanced based on the number of falls. The two classes fall and no falls each comprised of 1,621 results each for both wave one and wave two to ensure an accurate result was produced in both waves.

Classification via Regression was not the only model used. See Table 3 and 4 for the list of models and results that were all performed on the TILDA dataset in wave one and two. There are no drastic differences in the model results, however the models that did perform the best at predicting falls are the ones highlighted in green. In wave one the highest model, Classification via Regression classifier predicted 62% and in wave two there were three models, Decision Tree, Simple Logistic and Classification via Regression that performed the highest with a percentage of 69.

Table 1 Weka Results for Wave 1

Table 2 Weka Results for Wave 2

When running each of the models, each input risk factor was added incrementally one by one to establish whether or not the risk factor had an effect on falls. This took a very simplistic approach to distinguish which risk factors were of more importance and build upon each risk factor.

Figure 3 presents a Receiver Operating Characteristic (ROC) curve result, which is created by plotting the true positive rate against the false positive rate at different thresholds, for the Classification via Regression example in wave two with an

Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.696.

Figure 2 ROC Curve Plotted for Classification via Regression

Figure 4 presents a ROC curve result, for the decision tree example in wave two with an AUROC value of 0.579 for falls and 0.799 for no falls.

Weka Classifier Correctly Classified %Naïve Bayes 0.61SMO 0.60PART 0.60Random Forest 0.57Decision Tree 0.59Bayes Net 0.61Logistic 0.60Multilayer Perceptron 0.56SGD 0.59Simple Logistic 0.60Classification via Regression 0.62

Weka Classifier Correctly Classified %Naïve Bayes 0.63SMO 0.64PART 0.67Random Forest 0.68Decision Tree 0.69Bayes Net 0.63Logistic 0.67Multilayer Perceptron 0.67SGD 0.66Simple Logistic 0.69Classification via Regression 0.69

Figure 3 ROC Curve Plotted for a Decision Tree

Figure 5 presents a ROC curve result, for the logistic example in wave two with an AUROC value of 0.886 for falls and 0.236 for no falls.

Figure 4 ROC Curve Plotted for a Logistic Model

IV. CONCLUSION

To conclude, the present study presented an association between health care risks in older adults with falling and possible recurrent falls in older adults in Ireland. The objective was to explore different machine learning algorithms with the TILDA data and produce a model with the best result for predicting falls in older adults. This study has successfully displayed models which are higher in accuracy than others with Classification via Regression producing the best result in both waves of data.

Further work on this study will explore more of the health risk factors associated with older adults and those with dementia to find out which ones are having more of a significant impact on falls. The English Longitudinal Study of Ageing (ELSA), could also be reviewed alongside TILDA to analyze which risk factors could be correlated between both samples and use within a machine learning algorithm to provide a larger sample size and possibly further the current results.

REFERENCES [1] Health Service Executive (2018) Conditions & Treatments - Falls.

Available at: https://www.hse.ie/eng/health/az/ (Accessed: 04/12/2018).[2] Stevenson, M., McDowell, M. and Taylor, B. (2017) Concepts for

communication about risk in dementia care: A review of the literature., . doi: 10.1177/1471301216647542.

[3] Allan, LM., Ballard, CG., Rowan, EN., Kenny, RA. (2009) Incidence and Prediction of Falls in Dementia: A Prospective Study in Older People, . doi: 10.1371/journal.pone.0005521.

[4] Oyebode, J. R., P. Bradley, and J. L. Allen. (2013) Relatives’ Experiences of Frontal-Variant Frontotemporal Dementia, , pp. 156–166. doi: /10.1177/1049732312466294.

[5] While, C., Duane, F., Beanland, C., Koch. (2013) Medication management: The perspectives of people with dementia and family carers. Doi.org/101177/1471301212444056

[6] Witt, de L., Ploeg, J. (2014) Caring for older people living alone with dementia: Healthcare professionals’ experiences, 15, pp. 221 - 238. doi: 10.1177/1471301214523280.

[7] Raab, Markus., & Gigerenzer, Gerd. (2015) The power of simplicity: a fast and frugal heuristics approach to performance science, . doi: 10.3389/fpsyg.2015.01672.

[8] Cruise, S.M., Hughes, J., Bennett, K., Kouvonen, A. and Kee, F. (2017) The impact of risk factors for coronary heart disease on related disability in older Irish adults. . doi: doi.org/10.1177/0898264317726242.

[9] Meyer, David. (2001). Support Vector Machines The Interface to libsvm in package e1071. R News. 1.

[10] JRHS 2018 – look at google drive [11] Designing a Machine Learning based Software Risk Assessment Model

– Look google drive [12] Bal, R. and Sharma, S. (May 2016) Review on Meta Classification

Algorithms using WEKA, .[13] Costa, B., Rutjes, A., Mendy, A., Freund-Heritage, R. and Vieira, E.

(July 17, 2012) Can Falls Risk Prediction Tools Correctly Identify Fall-Prone Elderly Rehabilitation Inpatients? A Systematic Review and Meta-Analysis, . doi: doi.org/10.1371/journal.pone.0041061.

[14] Rafiq, M., McGovern, A., Jones, S., Harris, K., Tomson, C., Gallagher, H. and de Lusignan, S. (2014) Falls in the elderly were predicted opportunistically using a decision tree and systematically using a database-driven screening tool.

[15] Brownlee, J. (March 16 2016) Understand Machine Learning Algorithms -   Supervised and Unsupervised Machine Learning Algorithms. Available at: https://machinelearningmastery.com/supervised-and-unsupervised-machine-learning-algorithms/