It was an observational descriptive study with cross-sectional design. Then ANN was applied as a machine learning classifier for predictive modelling. The study was conducted at the service area of Bagbazar Urban Health and Training Centre (UHTC), Kolkata, West Bengal, India, between April 2016 to July 2016. It is also the urban field practice area of Department of Community Medicine, RG Kar Medical College and Hospital, Kolkata. All the elderly people living permanently at that slum during the study period were considered as the study population. As per records, available at Bagbazar UHTC, there were total 126 elderly people living there permanently during the study period. Individuals who were not able to comprehend the study schedule due to illness or dementia were excluded from the study. Those who expressed unwillingness to become a study participant were also excluded from the study.

Sample size was determined with finite population correction. The formula is n=Nz²pq/{d²(N-1)+z²pq} [15], here n=sample size, N= population size/total no of person in geriatric age group under the service area of Bagbazar UHTC i.e., 126, z=1.96 at 95% confidence level, p=prevalence of depression among elderly person (46.9%) [6], q=(1-p), and considering d= Absolute precision/margin of error (5%). Using the formula and above values, calculated sample size was 95. Considering non response rate as 10%, the final sample size was 105. Each of them was selected by simple random sampling without replacement technique using random number table. The selected elderly persons were interviewed at their home using pre designed and pre tested schedule after taking informed consent.

Ethical clearance was obtained from the Institutional Ethics Committee of RG Kar Medical College and Hospital, Kolkata, West Bengal, India.

A predesigned and pretested schedule was used for this study. It consisted of three parts. First part was sociodemographic information i.e., name, age, gender (Male/Female), living spouse (Yes/No), type of family (Nuclear/Joint), literacy (Literate /Illiterate), occupational status (presently working/ presently not working), personal income (Yes/No), and substance abuse (Yes/No); second part was information about presence or absence of co morbid conditions like hearing problem (Yes/No), visual problem (Yes/No), mobility problem (Yes/No) and sleeping problem (Yes/No); third part was 30 items Geriatric Depression Scale (GDS) [6] for assessment of depression (Normal/Depressed). GDS score 0-9 is normal and 10-30 is depressed [8].

Morbid conditions were assessed subjectively as per response made by the study participants. Hearing problem was considered to be present if there was any hearing difficulty with or without any hearing aids. Visual problem was considered to be present if there was any problem in visual acuity due to any reason after best possible correction. Mobility problem was considered to be present if there was difficulty in moving around for activity of daily living due to any reason. Sleeping problem was considered to be present if any problem related to falling asleep was reported by the study participants.

ANN predictive modelling, its performance analysis and development of a depressing prediction system was done using WEKA version 3.8.0 software. Weka is open source data mining software in the field of machine learning technology [9]. Primary data, collected by interviewing the selected 105 elderly people was used as a training data set for data mining in Weka. Then, the predictive model was build using ANN (Multilayer Perceptron, A feed forward ANN Model) as a classifier in the Weka. Performance of the build model was automatically analysed using 10-fold cross validation method.

Multilayer Perceptron (MLP), a feed forward ANN model was built in WEKA for classification with input variables i.e., age, gender (Male/Female), living spouse (Yes/No), type of family (Nuclear/Joint), Literacy (Literate/Illiterate), occupational status (presently working/ presently not working), personal income (Yes/No), and substance abuse (Yes/No), hearing problem (Yes/No), visual problem (Yes/No), mobility problem (Yes/No) and sleeping problem (Yes/No). The ANN model was designed with seven interconnected neurons in one hidden layer. The learning rate of the network was set at 0.3 for 500 iterations. Output was classified as normal i.e., non depressed and depressed.

A test data set were prepared selecting 30 geriatric individual by simple random sampling from the primary data set. Then this test data set were put into the ANN model in Weka with all input variables but output variable (Normal i.e., non depressed or depressed) were left blank. Then prediction was done by the ANN model which was previously trained with primary data set.

At the end of this research work a Microsoft Windows based application was designed using this trained ANN model to predict depression. This software was in its prototype version. This was easy to use; require very basic computer skill like data entry and computational time was less than one second.

The prevalence of depression among the study population was 45.7% at Bagbazar Slum in Kolkata. Mean (±standard error) age of the study participants were 66.6(±5.6) years. Among them 57.1% were male, 66.7% had living spouse, 73.3% were living in a joint family, 69.5% were literate, 58.1% had personal income, 54.3% were engaged in some kind of occupational activity, and 52.4% had a habit of substance abuse. Morbidity profile of the study population showed that 51.4% were suffering from visual impairment, 54.2% suffering from mobility problem, 21.9% from hearing difficulty and 19.1% from sleeping disorder. A comparison between depressed and non-depressed or normal study participants with various sociodemographic [Table/Fig-1] and morbidity factors [Table/Fig-2] are tabulated.

ANN model was trained and tested on the primary data set with 10-fold cross validation method. Result showed that among 105 instances it correctly classified 102 (97.14%) instances and incorrectly classified instances 3 (2.86%) with a Kappa statistic 0.94, Mean absolute error 0.03, and root mean squared error 0.14. Accuracy of the model showed weighted precision 97.2%, True positive rate 97.1%, false positive rate 3.1%, recall 97.1%, F measure 0.97, MCC (Matthews Correlation Coefficient) 0.943 and ROC were 0.99.

Result showed that the trained ANN model correctly diagnosed all 30 instances (12 depressed and 18 non depressed) with a Kappa statistic 1.

Automated disease prediction system using artificial intelligence is an upcoming concept in health care delivery system. Diagnosis of depression using machine learning technology is one such application of artificial intelligence. A study conducted in India by Mukherjee S et al., showed that ANN model can diagnose depression from various somatic symptoms, with reasonable accuracy with nominal root mean squared error (0.11) [16]. In our study, we are able to predict depression from socio economic and somatic symptoms with 97.2 % accuracy and root mean squared error was 0.14. Another study conducted by Khalili M et al., at Iran concluded that ANN modelling approach is a better way to diagnose depression and mania with acceptable accuracy with minimal error (0.13) [17]. Study conducted by Tortajada S et al., in Spain predicted postpartum depression using ANN model with a reasonable accuracy (82%) [18]. A study, conducted by Bhakta I et al., to compare different machine learning classifiers for predicting depression among the senior citizens, found that multilayer perception, a feed forward ANN model, had the predicting accuracy of 87% [19]. In accordance with the published articles, our study also recognises the power of ANN model in prediction of depression, especially among the older individuals.

This study was carried out in only one slum area in Kolkata, so an ANN model based on primary data obtained from large cross -sectional studies involving various parts of the country would have been close to generalization. Moreover, this will help to built a robust prediction system.

This ANN model based automated geriatric depression prediction system will help health care providers, to identify those geriatric persons who are more at risk of being depressed. Moreover, prediction can be made within seconds using sociodemographic and morbidity factors easily. At risk individuals can be referred to consultant psychiatrist for farther follow up and management. This will help to decrease the burden of morbidity and mortality due to depression among the geriatric population.