Discovering risk patterns in people with affective disorder–induced disabilities associated with their healthcare delay

Abstract

Background

People with affective disorder–induced disabilities (ADIDs) often experience complex needs that delay their healthcare. Discovering hidden patterns in these people for real-world use of health services is essential to improve healthcare delivery.

Methods

A cross-sectional study population (2501 adults with ADIDs) was obtained from the Australian national representative survey of disability in 2015, including 21 demographic, health and social characteristics and healthcare delay information in general practice, specialist and hospital services. The Self-Organising Map Network was used to identify hidden risk patterns associated with healthcare delay and investigate potential predictors of class memberships by means of simple visualisations.

Results

While experiencing disability avoidance showed across different healthcare delays, labour force appeared not to have any influence. Approximately 30% delayed their healthcare to general practice services; these were young, single females in great need of psychosocial support and aids for personal activities. Those who delayed their healthcare commonly presented a lack of social connections and a need for contact with family or friends not living in the same household.

Conclusions

The pattern evidence provides an avenue to further develop integrated care strategies with better targeting of people with ADIDs, considering social participation challenges facing them, to improve health service utilisation.

Introduction

Mental disorder is one of leading causes of disabilities in terms of functional impairment, activity limitation and participation restriction¹ and continues to grow with substantial health, social and economic consequences around the world.² Affective (mood) disorders are common and comprise a range of mental health conditions changing how people feel and respond to emotions.³ Specifically, people with affective disorders may experience various forms of clinically significant mental struggles, of which herein we consider depression, anxiety and other stress-related psychoses and somatoform disorders (excluding postnatal depression and substance-induced mood disorder) that result in some form of disability.

Such affective disorders may persist over a long time with symptoms developing due in part to other comorbid conditions or getting worse in response to environmental catalysts such as social isolation, and therefore regular sessions of intervention therapies have been encouraged.⁴ Moreover, living with disability may perhaps further complicate the needs of therapy and support. Thus people with affective disorder–induced disabilities (ADIDs) are vulnerable while facing exacerbation of symptoms requiring urgent healthcare.⁵ Under circumstances of delayed healthcare, people with ADIDs experience serious illness, worsening prognosis and possible loss of life.⁶ Assessment of health service utilisation in people with ADIDs is essential for developing appropriate strategies and policies to improve their health and life.

Delivering optimal healthcare to people with ADIDs requires careful attention to their complex care needs to increase their ability to carry out day-to-day activities. However, people with functional and activity limitations experience many barriers to healthcare, especially in the absence of policy modification.⁷ Delays in healthcare vary across people with different characteristics.^6,7 Thus clarification of different health service utilisation may assist mental health professionals to optimise care delivery, communicate with patients and their families and create tailored care approaches. This in turn can reduce the ineffective deployment of mental health resources and improve healthcare delivery and outcomes.

The current literature indicates a dearth of studies attempting to understand comprehensive patterns of health service utilisation in people with ADIDs and discover knowledge on potential risk factors associated with delayed healthcare for decisions to improve mental health support in this vulnerable population.^5–7 In this study we aimed to identify patterns to real-world use of health services in people with ADIDs by characterising them with demographic, health and social factors associated with their healthcare delay. Our findings may provide reliable evidence and support evidence-informed decision making for healthcare providers in the development of appropriate mental healthcare plans and pathways.

Methods

This study employed a cross-sectional design for the analytic data and identified patterns of individual characteristics among people with ADIDs (input data) in association with their delayed health services (output data). To model the pattern analysis, we applied an advanced machine learning and visual analytics approach with supportive visual modelling techniques.

Study data

We obtained data from the 2015 Survey of Disability, Ageing and Carers (SDAC) in Australia,⁸ which was the most recently available at the time of analysis for this study. Using multistage random cluster sampling schemes, the Australian Bureau of Statistics (ABS) has conducted the SDAC since 1981, with strict quality control measures to ensure survey coverage, reliability and confidentiality. This involves interviewers trained to carry out individual surveys by screening for eligible candidates, interviewing those who agreed to participate and entering responses to a wide range of questions. Proxy interviews were also carried out for those having problems, such as illness or language barriers, answering questions for themselves.

The 2015 SDAC comprises holistic information of participants’ need for and receipt of assistance across common activities of daily living. Additional efforts were made to improve responses by the use of follow-ups and/or proxy interviews. All respondents provided consent to participate in the survey. As a nationally representative sample of people with disabilities, the 2015 SDAC population was comprised of 63 515 non-institutionalised respondents, with a response rate of 80% in private dwellings and 90% in self-care retirement villages, and an additional 11 696 institutionalised respondents with a response rate of 90%. In this study we considered non-institutionalised respondents ≥18 y of age with determined ADIDs as the study population. Based on the agreement between the ABS and Universities Australia, confidentialised microdata were made available, and we retrieved the information specified for people with ADIDs (2501 Australians) on their health conditions and needs as well as their sociodemographic backgrounds. Ethics approval for the ABS to conduct the household interviews was granted under the Census and Statistics Act 1905. Additional approval was granted by the Australian National University Science and Medical Delegated Ethics Review Committee (2017/175).

Delayed healthcare (output data)

We identified self-reported ever-delayed use of health services in the last 12 months by three major types: general practice (GP), specialist (SP) and hospital (HP) services. A general association in terms of severity of illness or condition exists for these three different forms of regular care, with GPs providing primary and preventive care as gatekeepers, SPs providing for more serious conditions and/or offering more complex treatments and HP providing for conditions serious enough for inpatient care.

According to the ‘yes/no’ dichotomous response to SDAC questions (e.g. ‘Has there been any time [name] needed to see a GP but delayed going?’), the value of healthcare delay to each service was assigned as binary, ‘yes’ (1) or ‘no’ (0) for delayed use or not, respectively. The three service types were formed as the output dataset (‘Healthcare Delay’) in this study. Although they are mutually exclusive health services, people with ADIDs might have complex healthcare needs for comorbid and/or underlying health conditions and use multiple types of the services during a certain period of time. Therefore we also considered different combinations of service delay in the real world.

People characteristics (input data)

Individual characteristics of personal healthcare needs were grouped into three broad categories: demographic, health and social factors. Table 1 describes a total of 21 factors with their index and values specified for this study, forming the input dataset (‘Demo-Health-Socio Factor’).

Table 1.

Number and proportion of people by each value of 21 input factors^a

Demographic factors			Health factors			Social factors
	n	%		n	%		n	%
D1. Age group (years)			H1. Long-term treatment			S1. Number of home-away activities
15–44	684	27.35	No	534	21.35	None	231	9.24
45–64	1126	45.02	Yes	1967	78.65	1	516	20.63
≥65	691	27.63	H2. Severity of impairment			2	641	25.63
D2. Ruralityb			No	501	20.03	3	509	20.35
Metro	1529	61.14	Mild	799	31.95	4	331	13.23
Inner	575	22.99	Moderate	463	18.51	5	157	6.28
Outer	397	15.87	Severe	478	19.11	≥6	116	4.64
D3. Sex			Profound	260	10.40	S2. Internet use
Male	947	37.86	H3. Number of chronic conditions			No	1774	70.93
Female	1554	62.14	1	110	4.40	Yes	727	29.07
D4. Country of Birth			2	340	13.59	S3. Labour force
Australia	1928	77.09	3	418	16.71	No	1842	73.65
MESCc	264	10.56	4	420	16.79	Yes	659	26.35
NESCd	309	12.36	5	334	13.35	S4. Leave home as wish
D5. Single status			6	271	10.84	No	1081	43.22
Single	1572	62.85	7	210	8.40	Yes	1420	56.78
Partnered	929	37.15	≥8	398	15.91	S5. Social attendance
D6. Education			H4. Comorbid physical impairment			No	156	6.24
No	1733	69.29	No	696	27.83	Alone	1072	42.86
Yese	768	30.71	Yes	1805	72.17	Accompanied	1273	50.90
D7. Concessionf			H5. Comorbid psychosocial impairment			S6. Want more visitsh
No	701	28.03	No	1172	46.86	No	1449	57.94
Yes	1800	71.97	Yes	1329	53.14	Yes	1052	42.06
D8. Housing tenure			H6. Personal aidg			S7. Disability avoidance
No Mortgage	788	31.51	None	971	38.82	No	1102	44.06
Part Mortgage	539	21.55	Single	841	33.63	Yes	1399	55.94
Rental	1174	46.94	Multiple	689	27.55

^aThe study population (n=2051) and the factor values were normalized in the order from top (e.g. no=0) to bottom (e.g. yes=1) for the pattern analysis.

^bRemoteness of residence using the ARIA+.²⁹

^cOther English-speaking countries.

^dNon-English-speaking countries.

^eCompletion of at least year 12 courses.

^fBenefits having access to public (e.g. healthcare) services at a lower cost.

^gNumber of personal activities for which aid needed due to disability.

^hWanting more contact with family/friends not living in the same household.

Demographic factors included (D1) age groups categorised as 18–44, 45–64 or ≥65 y; (D2) remoteness of residence using the Accessibility/Remoteness Index of Australia (ARIA+) as major cities, inner regional or remote areas; (D3) sex as male or female; (D4) country of birth as Australia, other mainly English-speaking countries (MESC), including the USA, UK, Canada, Ireland, South Africa and New Zealand, or non-English-speaking countries (NESC); (D5) marital status as single or partnered; (D6) educational attainment for completion of at least year 12 as yes or no; (D7) possession of government concession cards with benefits providing access to medicines and healthcare services at a lower cost as yes or no; and (D8) housing tenure as owner without a mortgage, owner with a mortgage or renting.

Health factors included (H1) having treatment for long-term conditions or not; (H2) severity of impairment as having no, mild, moderate, severe or profound limitations in core activities; (H3) number of chronic conditions as 0, 1, 2, 3, 4, 5, 6, 7 or ≥8; (H4) comorbid physical impairment as yes or no; (H5) comorbid psychosocial impairment as yes or no; and (H6) number of personal activities for which aid is needed due to disability as none, single or multiple.

Social factors included (S1) the number of social activities participated in away from home, categorised as 0, 1, 2, 3, 4, 5 or ≥6; (S2) having digital activities using the internet as yes or no; (S3) participating in the labour force as yes or no; (S4) leaving home as often as you would like as yes or no; (S5) attendance at main cultural, sport and social activities as none, alone or accompanied; (S6) wanting more contact with family/friends not living in same household as yes or no; and (S7) avoid situations due to disability as yes or no.

Data analysis

In this digital era, acquisition of an enormous amount of data in the healthcare sector has increased the data complexity and uncertainty with multivariability, high-dimensionality, non-linearity and non-classification. Consequently, there exists a growing need for advanced methods to explore the data patterns and discover knowledge through evidence-driven decision-making processes that are difficult/impossible for conventional statistical and visualisation approaches.^9,10 The current study data (2501 observations with 21 input variables and 3 output variables) were multivariate, high-dimensional, non-linear and unlabelled in nature. In order to explore evidential and insightful pattern information from such data, we applied a machine learning–based visual analytics approach to the pattern recognition in this study.

Self-Organising Map Network (SOMNet) model

SOM is an artificial neural network, an unsupervised machine learning algorithm, developed to explore patterns of complex data that have no prior knowledge (e.g. class information of data).¹¹ SOM learns an original dataset based on data similarity and visualises the pattern information on an abstract map (e.g. hexagonal grid map). The grid of a SOM is composed of a set of neurons that are topologically organised and adaptively represent a data space by reflecting the data properties.

Using a SOM, the pattern of a single dataset (either input or output dataset) can be explored. Based on its machine learning and information visualisation capabilities, the SOMNet was developed to explore associative patterns between different datasets (input and output datasets) by utilising multiple SOMs in a network fashion.^12,13 The SOMNet incorporates various visualisation techniques that enable users (e.g. healthcare experts) to perform pattern exploration within a dataset and between datasets by visually processing and interpreting analytical information. The SOMNet has been applied to complex case studies, including the mental health domain for evidence-informed healthcare policy planning.¹⁴

Using the SOMNet, we modelled the data association by networking the input dataset (Demo-Health-Socio Factor) and the output dataset (Healthcare Delay). Based on the functional design of the SOMNet, the main goals defined for its application in this study are (1) input pattern identification of people characteristics, (2) output pattern identification of their delayed healthcare and (3) associative pattern identification between (1) and (2).

Supportive visualisation

The SOMNet allows users to compare the differences in characteristics of people with ADIDs using its property shape visualisation. However, it can be less friendly and difficult for lay users (e.g. healthcare practitioners) to capture meaningful information without training.^10,14 In this study we invented simple visual modelling techniques to present the SOMNet results of the complex patterns and support users in their understanding of the pattern information and interpreting their findings.

A feature visualisation in Figure 1 was introduced to present the SOMNet pattern results of the 21-dimensional input factors/variables. Each part of the feature was allocated with one of the input factors. Based on the order of values for each factor in Table 1, the value was defined as low (top) to high (bottom), and the higher value range was defined for the feature range (bracket per factor index in Figure 1a). Hence the SOMNet pattern value of each input variable was depicted as binary, with grey for ‘yes’ and white for ‘no’ according to its value falling in the feature range (Figure 1b). The feature visualisation can facilitate comparison of the overall characteristics pattern of people with ADIDs and the factor association with their healthcare delay.

Figure 1.

A feature visualisation. (a) The 21 input factors/variables in the feature parts with their feature value range in brackets. (b) The pattern value in grey (‘yes’) if it is in the feature range or in white (‘no’).

Results

Of the study population (n=2501), there were more adults ≥45 y of age (D1, 72.6%), females (D3, 62.1%), singles (D5, 62.9%), born in Australia (D4, 77.1%), living in metro areas (D2, 61.1%), housing with mortgage or renting (D8, 68.5%), having concession cards (D7, 72.0%) and not completed year 12 courses (D6, 69.3%), as seen in Table 1. In relation to delayed healthcare, 30.3% (n=757) delayed their visits to the GP, followed by 18.4% (n=460) and 6.8% (n=169) delaying their SP and HP care, respectively (refer to Table 2).

Table 2.

Proportion of the input factors with specified values across healthcare delays

Type of healthcare delays		No (n=1560)	GP (n=757)	SP (n=460)	HP (n=169)	All (n=81)
Demographic factors	Index	%
Older adults (≥45 y)	D1	80.51	57.07	57.40	53.85	45.68
Remote residents	D2	15.96	14.53	13.70	11.24	3.70
Females	D3	60.77	64.33	66.30	68.64	66.67
Non-English-speaking background	D4	13.46	10.57	10.43	9.47	6.17
Single	D5	58.72	71.60	71.96	73.37	79.01
Completion of year 12	D6	28.53	34.61	38.04	39.05	48.15
Concession card holders	D7	72.12	70.94	71.74	77.51	75.31
Renting	D8	41.67	55.88	59.13	60.36	64.20
Health factors	Index	%
Treatment for long-term conditions	H1	79.17	76.49	80.00	87.57	87.65
Severe–profound limitations	H2	27.56	32.63	31.74	49.12	43.21
Comorbid chronic conditions (>2)	H3	80.39	83.89	85.00	89.94	87.64
Physical impairment	H4	70.38	74.77	76.09	83.43	76.54
Psychosocial impairment	H5	47.12	64.60	64.57	75.15	82.72
Multiple personal aids needed	H6	25.19	32.10	31.09	49.70	46.91
Social factors	Index	%
No home-away activities	S1	9.10	9.25	10.22	13.02	18.52
Internet user	S2	34.62	18.23	17.61	24.26	13.58
Employed in the labour force	S3	23.65	33.03	31.52	18.34	19.75
Leave home as wish	S4	65.32	40.03	43.26	30.18	25.93
Social participation alone	S5	45.26	39.37	35.43	38.46	35.80
Want more visits	S6	36.15	52.18	55.87	60.36	65.43
Disability avoidance	S7	45.77	75.03	74.13	81.66	90.12

Individual pattern identification of input and output

The input data space was projected on the input SOM as shown in Figure 2. A total of 13 different pattern clusters (C1–C13) were identified for the study population of 2501 people with ADIDs (Figure 2a). Each cluster was explained using a shaped radar graph (property shape; Figure 2b) that shows the values distribution of the 21 demographic, health and social factors. Different clusters show different property shapes for the 21 input variables (e.g. circled clusters, C1 and C12 in red and blue, respectively).

Figure 2.

The Demo-Health-Socio Factor patterns on the input SOM. (a) Cluster numbers in different colours randomly generated. (b) Data property shapes of (risk) factors, connecting each line from the centre (low) to the edge (high) and marking each variable index clockwise.

Figure 3 describes the pattern information of people characteristics approximated for every cluster on the input SOM. For instance, the input cluster C1 was comprised of people with ADIDs who were young (D1), living in metro areas (D2), female (D3), born in Australia (D4), partnered (D5), having a higher educational attainment (D6), without a concession card (D7) and having a mortgage (D8). They had long-term healthcare plans (H1), non-severe impairment (H2), fewer chronic conditions (H3) and personal aids (H6), and comorbid physical (H4) but no psychosocial conditions (H5). They were participating in the labour force (S3), not relying on internet use (S2), having social activities (S1, S4 and S5) and likely staying socially connected (S6 and S7). In comparison with C1, the input cluster C12 grouped people with ADIDs having a lower educational attainment (D6) with concession cards (D7), comorbid psychosocial conditions (H5) and more personal aids (H6). In stark contrast to C1, the people in C12 were unlikely to participate in the labour force (S3), rely on internet use (S2), have social activities on their own (S1, S4 and S5) and stay socially connected (S6 and S7).

Figure 3.

A heat-map for the Demo-Health-Socio Factor patterns of the input SOM. Four clusters (C1, C6, C8 and C12) are exampled in the red box for the associative pattern identification (the darker the colour, the higher the probable value in the feature range of each input index in Figure 1a).

The output data space is projected on the output SOM in Figure 4. A total of eight different pattern clusters were identified for the study population based on their delayed healthcare to the three different health services/variables (Figure 4a). Some cluster numbers with only a few neurons on the map were joined into the most similar neighbour cluster if their property shape showed a very small difference (Figure 4b). Thus the eight clusters identified for the output SOM were C1, C3, C5, C8 (including 7, 10 and 17), C9 (including 4 and 6), C13 (including 2, 11 and 12), C15 (including 14 and 16) and C18.

Figure 4.

The Healthcare Delay patterns on the output SOM. (a) Cluster numbers in different colours are randomly generated. (b) Data property shapes of healthcare delay, connecting each line from the centre (low or no delay) to the edge (high or delay) and marking each variable index clockwise.

As the output SOM was trained using only three variables, the property shape was much clearer for the pattern identification (e.g. circled output clusters, C1 and C3 in red and blue, respectively). The output cluster C1 was comprised of people with ADIDs who had ever delayed GP service only, while people in the cluster C3 showed the pattern of having healthcare delay of both GP and SP services. The output SOM revealed all possible patterns of healthcare delays in people with ADIDs with respect to their health service utilisation in the real world.

Associative pattern identification between input and output

From the input SOM, we examined four clusters (C1, C6, C8 and C12) of people characteristics to identify how different they are and how they are associated with healthcare delay on the output SOM (input-driven analysis). For better understanding of the input differences, Figure 5a presents the people characteristics patterns of the four clusters in accordance with their feature visualisations. Given each of the input patterns in Figure 5a, the highly weighted (darker colour with white dot for >50% weight association) output regions (associative patterns) in each output SOM were explored through the SOMNet analysis as shown in Figure 5b. We also summarised each of the associated healthcare delay patterns in Figure 5b into each corresponding triple hexagon visualisation for the three health services (see Figure 5c).

Figure 5.

Visualisations of the input-driven associative output patterns. (a) The feature visualisation of four input patterns (C1, C6, C8 and C12) of people characteristics. (b) The output SOM patterns of healthcare delay associated with the input patterns in (a). (c) The triple hexagon visualisation for the output pattern information in (b).

Figure 5 presents the visual information for identification of different patterns of healthcare delay in people with ADIDs having different characteristics. For instance, the input clusters C6 and C8 had their great health needs (H1–H6) for care in common. However, people in C6 were mostly young (D1) and single (D5) females (D3), demonstrating a distinctive pattern with likely delaying their GP and SP services, in contrast to their counterparts in C8 who did not have any healthcare delays.

Conversely, associative patterns of the input for some given outputs were also identified using the SOMNet model (output-driven analysis). As shown in Figure 6, three healthcare delay patterns—O1 (no delay), O2 (GP only delay) and O3 (all delays)—were given to the output SOM (Figure 6a) and the associated weights were highlighted in the input SOM (Figure 6b). The associative people characteristics property of the highly weighted input SOM region for each given healthcare delay output pattern was summarised in the feature visualisation (Figure 6c).

Figure 6.

Visualisations of the output-driven associative input patterns. (a) The triple hexagon visualisation of three output patterns (O1, O2 and O3) of healthcare delay. (b) The input SOM patterns of people characteristics associated with the output patterns in (a). (c) The feature visualisation for the input pattern information in (b).

Those who delayed their GP services (O2 in Figure 6), for instance, were more likely to be young (D1) and single (D5) females (D3), having concession cards (D7), living in rental accommodations (D8). Although they were likely attending some social activities (S5), they were experiencing avoidance due to disability (S7), experiencing a lack of social connections (S1, S2 and S4) and wanting more contact with family/friends not living in the same household (S6). This group had fewer long-term medical conditions (H1, H2 and H3), but was in great need of personal aids for daily activities (H4 and H6), including cognitive and emotional tasks (H5).

When comparing the feature visualisations with the triple hexagon visualisations for the pattern results in the input- and output-driven analyses, we could validate the association by finding the common risk patterns in people with ADIDs and their healthcare delays. Clear variation in people characteristics was observed across different healthcare delays. People with ADIDs who were <45 y of age (D1), single (D5), female (D3) and living in rental accommodations (D8) (O2 and O3 in Figure 6c) were overrepresented in comparison with their counterparts who did not have any healthcare delays (O1 in Figure 6c). There seemed to be no outstanding health-related issues that might potentially impact on any healthcare delays, except that having comorbid psychosocial impairment (H5) might associate with healthcare delays. All those who delayed their healthcare demonstrated a lack of social connections (S1, S2 and S5), and in particular a need of contact with family/friends not living in the same household (S6). The majority having healthcare delays did not leave their home as often as they would like (S4). While experiencing disability avoidance situations (S7) was a common driver across different types of healthcare delays, exclusion from the labour force (S3) appeared to have no influence on delaying visits to health services.

Discussion

This population-based study focusing on health service utilisation in people with ADIDs demonstrated varying characteristics of those people that would impose challenges in the provision of appropriate and timely care to meet their complex needs. Although improving healthcare delivery in people with mental disabilities is a high-priority research area in Australia and internationally, there is limited understanding of delayed healthcare and how it varies across population groups,^5,6 calling for investment to achieve better targeting. This study used the advanced SOMNet approach and presented the first pattern recognition of complex healthcare needs in people with ADIDs, discovering hidden risk patterns associated with their healthcare delay in unique ways. The risk factor profiles in association with delayed healthcare varied to different extents across the study population, highlighting their need for social connections. These findings could assist mental health professionals in distinguishing certain groups of people with ADIDs at risk of delayed healthcare and develop countermeasures to improve their health service utilisation.

The risk factors indicated their unique and combined roles of differentiating cluster membership. As healthcare needs in people with ADIDs are complex,^5,6 their expectations towards healthcare provision are not homogeneous. It has become fundamental to provide collaborative personalised services and avoid delays in healthcare.^15–17 There are multiple possible views of the variation. For example, delayed healthcare was not that paramount in the younger group, except that they were more likely to delay GP services, whereas the prominent challenge of healthcare delivery was related to the delayed services in their older counterparts who might have been experiencing more severe limitations. However, such a proposition is speculative and requires future investigation to confirm its validity.

In spite of differentiating the risk factors, having treatments for long-term conditions and attendance at main cultural, sport and social activities were consistently observed in the study population. This finding indicated a continuous and integrated care process at play in the current setting, with the augmented approaches to encourage community participation.¹⁸ Intervention strategies focused on making good use of social care alongside traditional health services would potentially meet the healthcare needs in people with ADIDs.

Delayed use of GP services followed by SP services was common in people with ADIDs who were young and single females, experiencing disability avoidance and in greater need of aids for personal activities and more visits from families or friends, in contrast with those who did not delay their use of GP services. No evidence was found that people having comorbid physical impairment would experience GP delays as a conventional assumption, especially when taking physical inaccessibility into account. Those who had comorbid physical impairment actually experienced little disability avoidance, implying unhelpful attitudes and behaviours rather than inaccessible physical environments might be at play for delayed use of GP services.^19,20 While it cannot be assumed that they may have fewer healthcare needs, the finding demonstrated the need for psychosocial support in this vulnerable population.

One possible view for delayed used of GP services would be related to unpleasant GP encounters in the past. GP services might not have been delivered with multidisciplinary support from properly trained mental health professionals. Therefore it is worth noting that the collaborative effort between GPs and mental health professionals to improve outcomes in people with ADIDs is an area warranting further investigation. Nonetheless, intervention strategies for people with ADIDs presenting primarily with multiple psychosocial disorders should involve personal aids for daily activities, including cognitive and emotional tasks to improve their utilisation of GP services.

People with ADIDs demonstrated social and psychological needs, which would be paramount for them to realise their potentials in everyday life. This study found those at risk of delayed healthcare would like to have their family and friends visit them. Such need may not be related to their physical health, but spiritual and psychological health instead, which was consistent with previous findings.^21,22 Conversations with close family or friends seems to be crucial to deliver the relief and comfort they need, which might differ from what healthcare professionals normally provide. Discomfort during care can come from various sources, including lengthy waiting times and discrimination by staff,^23–25 thus people with ADIDs might turn to their family and friends for help with practical tasks and emotional distress.

Although inclusion of family members or primary caregivers in the relevant decision-making process during healthcare encounters has been recommended,²⁶ this approach is practically difficult and inconsistently implemented across contexts. Hence successful interaction with people with ADIDs during their healthcare encounters might begin with an understanding that people with mental disorders are capable of participating in social activities and they do not have the same needs for care. This understanding may pave the way to reduce disability discrimination and ensure the same societal goal of equal opportunity of care.

Future directions and implications

Rather than using conventional models to estimate effects of risk factors in association with delays in healthcare, we designed with an intention to identify the complex patterns in people with ADIDs as an example to allow healthcare planners or providers to make better informed decisions using the unsupervised machine learning and visualisation techniques without the need to assume any underlying statistical distributions of the candidate risk factors. This is of vital importance in the evolving healthcare landscape, especially when facing people with complex needs of multisector efforts. The application of SOMNet demonstrated advantages in dealing with feature extraction and pattern discovery of high-dimensional data, while conventional regression modelling approaches would be swamped by the multicollinearity problem in this study. Hence investment should be encouraged to build capacities in the healthcare sector with respect to understanding, using and developing pattern recognition algorithms to enhance service planning and delivery.

Given the globally recognised right to high-quality healthcare for all people with mental disabilities,^5–7 it is imperative to tailor personalised programs for them that target the most relevant care patterns. Currently there are multiple programs aiming to improve the quality of healthcare.^27–29 This study adds to the supportive evidence related to these programs and emphasises the importance of identification of utilisation patterns to improve care delivery and outcomes. Future interrogation of additional administrative healthcare data with detailed therapeutic information,³⁰ as well as in-depth interviews with people with ADIDs and their primary caregivers, may further reveal cues for action to change service provisions. Nevertheless, the current pattern recognition combining the demographic, health and social characteristics of people with ADIDs can facilitate the establishment of appropriate integrated social and healthcare programs, thereby not only avoiding healthcare delays, but also improving patients’ social life as a global achievement. This reinforces the necessity to promote appropriate social rehabilitation services for people with ADIDs to improve their health and social outcomes.

Limitations

There are several limitations in this study. First, it is a descriptive study of observed patterns in a cross-sectional dataset and does not explain the underlying reasons for those observed patterns. It is possible that some people with ADIDs delayed their healthcare for undeclared reasons. Second, considering that affective disorders comprise different subtypes with variations in severity, the lack of clinical data limited the ability to comprehensively consider healthcare scenarios for risk adjustment. It would be prudent to be cautious extrapolating beyond the current study setting. Third, delay in healthcare service utilisation was surveyed using a dichotomous response (i.e. yes or no), without differentiation according to the number of delays. Although a binary outcome reflected the proportional change of the study population with or without healthcare delay and appeared appropriate in the current pattern analysis, considering the number of delays would facilitate further population segmentation for better targeting. For example, future studies may be designed to unveil the difference between short and long or isolated and regular delay patterns. Fourth, the SDAC collected a substantial amount of self-reported data across different geographic regions in Australia during a relatively short period. However, with respect to affective disorders as a cause of disability, the respondents might lack clinical knowledge to justify the underlying causal mechanisms for their disability or impairment and therefore the current study perhaps did not capture all people with ADIDs. Moreover, people with ADIDs were rarely identified as having an indigenous background in this study, for reasons including that the current setting excluded those who live in cared accommodations, and the SDAC did not cover discrete aboriginal communities from very remote areas. Considering such limitations about self-reported data and underrepresented indigenous people, results should be interpreted with care. Lastly, but not least, considering statistical variation of the input and output data given the study population, the SOMNet model indicated some errors of associative sampling and recognition in group assignment. Thus the association results should be interpreted with caution. Nonetheless, the identified patterns of health service utilisation seem to be relevant and informative for those seeking to improve healthcare for people with ADIDs. Future enhanced human–machine collaborations may further facilitate such decision-making augmentation for health service planners and providers to stay abreast of the booming machine learning technology in the healthcare sector and build their capacities to easily grasp the concept of pattern classification algorithms and membership of the clusters and design interventions that match the needs of people in those clusters.

Conclusions

This study confirmed the heterogeneity and complexity of healthcare needs as well as their relationship to social care needs in people with affective disorder–induced disabilities. Recognition of diverse patterns of health service utilisation and identification of risk factors associated with healthcare delays provide an avenue to develop evidence-informed interventional strategies with better targeting to improve mental health service utilisation and healthcare outcomes they deserve. Concerted action is required by mental health advocates and policymakers to assemble this evidence, considering the social participation challenges facing people with ADIDs.

Authors’ contributions

WD and YC contributed equally to the manuscript, making substantial contributions to study conception, design and analysis, interpretation of data and revising the manuscript critically for important intellectual content. WD made a substantial contribution to data acquisition and interpretation. YC was a major contributor in the study design, data analysis, results interpretation and the first draft. Both authors read and approved the final manuscript.

Funding

This study was supported by the Department of Education of China (1125000172) and the Fundamental Research Funds for the Central Universities (3225002002A1).

Competing interests: None declared.

Ethical approval: Approval for use of the 2015 SDAC data was granted by the Australian National University Science and Medical Delegated Ethics Review Committee (2017/175).

Data availability

The data underlying this article are available from the ABS. The datasets were derived from the 2015 SDCA available from http://abs.gov.au/statistics/health/disability/disability-ageing-and-carers-australia-first-results/latest-release.