Abstract
Background:
The therapeutic paradigm in Alzheimer’s disease (AD) is shifting from symptoms management toward prevention goals. Secondary prevention requires the identification of individuals without clinical symptoms, yet “at-risk” of developing AD dementia in the future, and thus, the use of predictive modeling.
Objective:
The objective of this study was to review the ethical concerns and social implications generated by this new approach.
Methods:
We conducted a systematic literature review in Medline, Embase, PsycInfo, and Scopus, and complemented it with a gray literature search between March and July 2018. Then we analyzed data qualitatively using a thematic analysis technique.
Results:
We identified thirty-one ethical issues and social concerns corresponding to eight ethical principles: (i) respect for autonomy, (ii) beneficence, (iii) non-maleficence, (iv) equality, justice, and diversity, (v) identity and stigma, (vi) privacy, (vii) accountability, transparency, and professionalism, and (viii) uncertainty avoidance. Much of the literature sees the discovery of disease-modifying treatment as a necessary and sufficient condition to justify AD risk assessment, overlooking future challenges in providing equitable access to it, establishing long-term treatment outcomes and social consequences of this approach, e.g., medicalization. The ethical/social issues associated specifically with predictive models, such as the adequate predictive power and reliability, infrastructural requirements, data privacy, potential for personalized medicine in AD, and limiting access to future AD treatment based on risk stratification, were covered scarcely.
Conclusion:
The ethical discussion needs to advance to reflect recent scientific developments and guide clinical practice now and in the future, so that necessary safeguards are implemented for large-scale AD secondary prevention.
Keywords
INTRODUCTION
Alzheimer’s disease (AD) is the cause of 70% of all dementias [1], characterized by the combination of cognitive, behavioral, and functional decline, leading to loss of autonomy. AD represents a significant public health challenge worldwide. The disease course is understood as a continuum from the preclinical stage without cognitive symptoms, to mild cognitive impairment due to AD (MCI) and then dementia due to AD. Knowledge of the pathophysiology of AD has improved over the last decade, bringing about a deeper (albeit not conclusive) knowledge of genetic predisposition, identification of biomarkers (e.g., amyloid-β (Aβ) plaques in the brain or tau protein in cerebrospinal fluid) as well as new insights about their interaction with protective or disease-promoting factors [2–4]. In turn, these genetic, molecular, and environmental risk factors, or the subjective perception of declining cognitive capacities have been found useful to identify cognitively unimpaired individuals at higher risk of developing MCI due to AD and later on dementia due to AD [5–7]. Consequently, the therapeutic paradigm of AD has recently shifted from symptoms management in individuals diagnosed with MCI or dementia based on their clinical symptoms, to secondary prevention goals targeting “at-risk” individuals and aiming at modifying the natural course of the disease. Contributing to this shift are recent drug development programs testing earlier in the disease course compounds that previously failed in clinical trials (RCTs) on participants with MCI or dementia [8], in a hope that they can be efficacious if used earlier, even at preclinical stages of AD [9]. Recent claims that aducanumab, an anti-Aβ immunotherapy, improves cognition in patients with MCI or mild AD, lend some credibility to this approach.
In the research setting, participants are enrolled to clinical trials testing preventive treatments only if they have elevated AD biomarkers or genetic predispositions (cf. clinicaltirals.gov identifiers NCT02008357, NCT01931566). However, even such patients have a fairly low probability of developing AD in the future. Predictive modeling can help to produce a more accurate assessment of the probability of conversion from being cognitively unimpaired to MCI or dementia within a certain timeframe. Data used in such models: individuals’ demographics (e.g., age, sex, level of education), genetic markers (e.g., APOE4), and comorbidities (e.g., cardiovascular diseases) as well as longitudinally captured brain imagining metrics (e.g., PET scans to establish Aβ status) and results of cognitive tests, can typically be found in clinical registries from memory clinics. Even though as of today the applications of such models are mostly in the research setting, some clinics offer AD biomarker testing to their patients, followed by non-pharmaceutical intervention, e.g., lifestyle changes, cognitive rehabilitation, etc. In a future, aspirational scenario, predictive modeling could be applied in combination with a preventive treatment (currently not available), e.g., to identify patients with high risk of developing clinical symptoms of AD or patients likely to benefit most from the therapy.
A predictive model could also be developed based on minimal sets of demographic and clinical information. Such model could be used in a hypothetical scenario for broad (e.g., population) screening aiming to crudely sift out from general population individuals who might have an increased risk of developing AD in the future, so that these individuals could undergo further investigation using brain imagining, and other biomarker or genetic analyses.
Such new therapeutic paradigm in AD raises numerous ethical concerns and may have various social implications. Some of these concerns are typical for preventive medicine in general, yet at the core of the problem is AD’s specific setting—the need to intervene years or even decades before the onset of any cognitive, behavioral, or functional decline [11] without a certainty that an individual would ever develop clinical symptoms of AD, while the long-term consequences of these interventions are not yet fully understood. Uncertainty about the long-term consequences of future preventive AD treatment is due to the long natural history of AD which makes it impossible to evaluate in clinical trials, currently lasting up to 5 years in AD, all its clinical consequences. Likewise, the clinical trials will not be sufficient to fully appreciate the long-term societal consequences of a preventive intervention. The limited length of follow-up, narrow choice of endpoints, and stringent inclusion criteria are also limiting factors from the perspective of evidence needed for pharmacoeconomic assessment and drug reimbursement. This is another context where predictive modeling can and likely will be applied to remedy the knowledge gaps, e.g., through models bridging between strictly clinical trial endpoints (like neuropsychological assessment) and societally relevant outcomes (like institutionalization). Finally, the entire discipline of predictive medicine enabled by the technological and computational developments in the recent decades raises further ethical concerns and social implications.
A lively scientific debate about the ethical aspects of recruitment of pre-symptomatic individuals to clinical trials and observational studies has already been taking place in the recent years [12]. As AD prevention efforts will need to target a large number of people in order to be impactful, this debate will intensify. As soon as an efficacious preventive treatment is developed, a sense of urgency will arise to provide disease-modifying treatment (DMT) [13] to aging populations, to prevent public health crisis and the associated soaring burden of care.
The objective of the present study was to systematically review and discuss the ethical concerns and social implications raised by the use of predictive modeling in the setting of secondary prevention of AD. We focused on the types of arguments with particular relevance for current and future, anticipated, or aspirational clinical practice.
Here, we defined secondary prevention as targeting people “at risk” of AD dementia with an intervention aiming to prevent or delay the onset of clinical symptoms [14, 15]; and predictive modeling as the use of data from multiple individual subjects in statistical models to identify the likelihood of future outcomes—including patient-level outcomes—based on historical data [10].
Our specific research questions were identified through a preliminary, targeted literature search [16] and include the following: What are the ethical concerns and social implications associated with Selection of individuals for assessment of the risk of developing clinical symptoms of AD via predictive modeling, from a general population or population with known risk factors? The disclosure of individual’s risk of developing AD clinical symptoms assessed using predictive modeling? Preconditioning of access to AD preventive treatment, based on the predictive modeling, e.g., by selecting patients at high risk (in a future, aspirational scenario)? Assessment of the benefit-to-risk from AD preventive treatment administered at the preclinical stage, made using predictive modeling? What are the broader, population-level ethical concerns, and social implications of using predictive modeling tools in the setting of secondary AD prevention?
MATERIALS AND METHODS
Definitions
Whenever we refer to MCI or dementia we mean MCI due to AD, and dementia due to AD. The term “at risk of AD” refers here to being cognitively unimpaired but having an elevated risk of developing clinical symptoms of AD in the future, regardless of how this elevated risk was established (e.g., using genetic or biomarker analysis, or using an aggregation of risk factors from multiple data domains). “Preclinical AD” refers to cognitively unimpaired individuals with established AD biomarker. Whenever we use the term “preventive treatment”, we mean the future, aspirational drug targeting AD, used before AD clinical symptoms are developed.
Protocol development
The study protocol was prepared according to the reporting guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analysis for Protocols 2015 (PRISMA-P) [17, 18], registered with the PROSPERO international prospective register of systematic reviews (registration number CRD42018092205) on April 6, 2018 and published [19]. The completed PRISMA-P checklist is provided in the Supplementary Table 1.
Search methods
A comprehensive, systematic literature search was conducted between May and July 2018. The literature was retrieved from the Embase/Medline Daily, Scopus, and PsycINFO between 28 and 31 of May 2018 including coverage from 2007 until the search date. Additionally, a gray literature search was performed within pre-defined websites of relevant non-governmental organizations and professional associations, and using a generic Google search engine, where the first 10 pages of results were reviewed for potentially relevant entries. The full electronic search strategy is provided in the Supplementary Table 2.
Study selection
The systematic literature search followed the SPICE framework (Setting, Perspective, Intervention, Comparison, Evaluation) [20]. Included in the analysis were studies discussing ethical concerns or social implications, both from individual and societal perspective, both of using predictive modeling methods (statistical algorithms) and source data (e.g., demographics, genetic data, imaging data, cerebrospinal fluid examination, etc.) as a component of secondary AD prevention. Studies were included if discussing preclinical AD, including those with subjective memory complaint/cognitive impairment but without MCI diagnosis. We included studies reporting on the results of research on humans (basic, clinical, social, reviews/meta-analyses, observational, randomized controlled trials), including conference abstracts, editorials, commentaries, guidelines, discussion and position papers, books, and book chapters published in English, French, or German from the year 2007 onwards. The choice of this time span reflects the fact that secondary prevention is a recent therapeutic strategy against AD. Details of the study selection criteria are presented in Table 1.
Criteria for study selection with instructions for reviewers and examples
The retrieved abstracts were independently assessed by two reviewers and disagreements were adjudicated by a third reviewer. Reviewers had a possibility to exclude not eligible studies based on a review of the full-text versions prior to the extraction process.
Data extraction
Data were extracted from the eligible studies by single reviewers. The extraction was performed using a semi-structured extraction sheet where textual content extracted by all reviewers was uploaded in real time into an online spreadsheet for further qualitative data analysis. Text fragments were extracted according to the pre-specified research questions, aforementioned in the introduction of this paper, with a checklist of ethical concerns or social implications known to appear in this context. This checklist was derived from a seed of four studies [21–24] selected for this purpose by one reviewer and a bioethicist independent to this study. Lastly, for each of the research questions, open-ended text boxes were added to allow for capturing all novel themes, arguments, and considerations that were not initially included in the structured checklist. The extraction sheet development process is described in more detail in the systematic review protocol [19].
Data analysis
Extracted data were analyzed qualitatively using a thematic analysis approach [25, 26] defined as “a method for identifying, analyzing, and reporting patterns (themes) within data which minimally organizes and describes (your) data set in (rich) detail but also interprets various aspects of the research topic” [25]. The theme is defined as “a repeated pattern of meaning, capturing something important about the data in relation to the research question, and representing some level of patterned response or meaning within the data set” [25]. In characterizing salient ethical arguments, the focus was on the claims being made and the arguments supporting them, not on quantitative assessment of the number of times a given claim appears in the literature. Therefore, the frequency was not treated as a measure of importance. The research questions of this study defined the highest level themes which were further broken down into the lowest level of ethical and social considerations. In order to make sure that complex ethical arguments were understood in context, full-text papers were revisited during the iterative analytic process and reviewers were encouraged to use memos liberally during extraction and analysis. Both pre-specified and newly identified ethical and social considerations were then classified as either ethical concern or social implication, and grouped into themes. The connections and interdependencies between the themes were investigated. While the analysis relied upon self-nomination of ethical relevance by a reviewer, the grouping into ethical themes was matched with the additional mapping of the ethical principles establishing the perceived ethical relevance of each theme to the issue of predictive modeling. These principles were drawn from background literature in medical and public health ethics, including the four core principles of ‘principlism’ in biomedical ethics [27].
Risk of bias
One potential bias to a literature review is to treat what is most commonly reported as the most important. This bias is mitigated by the qualitative character of the present study, striving to understand a wide spectrum of the ethical and social concerns and disregarding their frequency in the literature. However, a potential inter-reviewer heterogeneity when different reviewers appraise manuscripts and documents in a different manner could result in some ethical arguments being missed or misinterpreted. To mitigate this bias, the team of reviewers participated in a face-to-face workshop on April 28, 2018 in Barcelona, during which the research objective, strategy, and extraction tools were thoroughly discussed and reviewed when needed. Further to that, the reviewers come from different backgrounds, including sociologist, clinical psychiatrist, psychologist, market access professionals specializing in AD, a pharmacist and market access professional, and a mathematician/statistical modeler. Finally, the results could be affected by a publication bias.
RESULTS
Study selection
The systematic literature search yielded in total 180 citations, 154 in bibliographic databases including Embase/Medline, PsycInfo, and SCOPUS and 26 through a manual search conducted in Google. After removal of duplicates, 152 abstracts were screened against the inclusion criteria and 92 were excluded at this stage. After full text screening, 14 additional publications were excluded. Reasons for exclusions are listed in Fig. 1. In total, 46 publications were retained.
PRISMA flow diagram.
Study characteristics
Of the forty-eight retained publications, thirty-two were journal articles with the majority of them coming either from medicine/gerontology (thirteen out of thirty-two) or interdisciplinary domain (twelve out of thirty-two). Four articles came from psychiatry or neuroscience field, and the remaining were published in social science or ethics journals. Seven further publications were either conference abstracts, proceedings, or presentations, six were reports, and three were books or book chapters.
Results of the individual studies
Table 2 summarizes the ethical concerns and social implication identified in the literature, structured along the research questions. Table 3 shows the matching of ethical concerns and social implication to one or more ethical principles.
Results of individual studies
*References are nested, meaning that a reference for a sub-theme populates the reference to the main theme whenever appropriate.
Mapping of ethical themes to underlying ethical principles
Selection of a population for risk assessment via predictive modeling
Who should have their AD risk assessed is one of the critical questions in the ethical debate around AD prevention. One approach could be population screening, e.g., screening everybody after a certain age, yet such an intervention might lead to “turning everyone into patients” [12, 28–33] and excessive operational burden for healthcare systems. Alternatively, model-based, precise assessment of AD risk could be made only among those with known risk factors for AD. While these two approaches belong to the classic arsenal of public health prevention, another unique concept identified in the literature was “screening whoever wants to be screened” [34], yet not without a question whether access to screening should be limited to individuals assessed beforehand as emotionally capable of eventually learning their risk status (e.g., not prone to depression). This concept is best summarized as “screening before screening” [12, 35–37]. Voluntary access to screening can be defended on pragmatic grounds by the fact that commercial genetic testing for AD is already available and will most likely come into large demand as soon as DMT is developed [34, 38]. Policymakers must ensure that healthcare and social systems are prepared in terms of implementation of laws safeguarding a growing number of patients, their data and their interests, and that professional and social policies are put in place to not only treat but also advise and educate them [28, 39–44].
Several ethical themes speak against assessing the risk of AD. The most prominent of them is the current lack of DMT rendering risk assessment not actionable [21, 45–53] and potentially even harmful, e.g., when side effects of invasive biomarker testing are considered [30, 41] or the threat of over diagnosis [4, 52] and competing risks are taken into account [37, 52]. The issue of competing risks is particularly valid in the AD setting, where at-risk or preclinical stage might span decades and where the elderly patient population might be prone to other age-related diseases. Further reservations against AD risk testing are: lack of adequate tests with sufficient predictive power to provide a trustworthy risk assessment [21, 52], lack of social consensus as to what predictive power could be considered sufficient [21, 38], and uncertainty, whether the presence of Aβ plaques is causally associated with AD [37, 42]. The latter argument is only partially relevant, though, in the predictive modeling setting, where co-occurrence can be sufficient to predict future outcomes.
Disclosure of individual risk assessed using predictive modeling
Considerations around disclosure practices do not differ substantially depending on whether they are based on genetic, biomarker, or imaging assessment, with an exception of the specific discussion on familial, early onset AD. Particularly relevant in this context of people with high risk are arguments in favor of disclosure due to the psychological benefits of resolving patient’s uncertainty of their AD risk [37, 51] and possibilities for future planning [4, 54–59]. Ethical considerations depend in turn to a large extent upon whether the disclosure is made in research setting in the absence of DMT versus in hypothetical clinical setting where DMT is available. In the latter case, there might even be an ethical obligation for disclosure. The governing ethical principle here is a postulate that the diagnosis should provide a patient with a benefit that overweighs the risks. Some papers, however, consider benefit much broader than access to treatment, pointing rather to the need of establishing whether a risk assessment brings clinically meaningful information [4, 31], considering patient’s individual situation, including the availability of support [39, 60] and their level of willingness to know their risk status, as these two might mediate the level of benefit from the diagnosis [4, 57].
On the other hand, major groups of arguments against disclosure address psychological harms associated with the remaining, post-testing uncertainty of the positive risk assessment until symptoms occur [12, 51] and even without certainty whether they will occur, given the possibility of a false-positive diagnosis [38, 51]. The ethical and social ramifications of a false-negative diagnosis are not specifically discussed in the literature. A very prominent theme in the literature stresses the risk of discrimination of people with high risk of developing AD symptoms within the workplace, healthcare system and society overall [21, 61] which might lead to their distress [4, 55] and potentially even objective, realistic limitation in how they perform in their daily life [12, 61]. AD risk assessment can bring about negative consequences not only to the patient, but to his or her relatives and significant others, as they might become anxious about their own risk [39, 40] or about the upcoming challenges of taking the role of a supporter or carer [39, 40].
Accurately communicating AD risk assessment to patients is considered challenging given the complexity of the issue, the differences between patients [28, 51] (e.g., in terms of their level of understanding of the disease and the uncertainty of preclinical risk assessment), level of agency and support, individual predispositions for depression; as well as unique statistical properties of particular methods which are used to make such a prognosis [38, 57].
Treatment: Preconditioning access to treatment and assessing benefit from it using predictive modeling
The relationship between the access to screening and to the treatment is reciprocal, meaning that the recommendation for screening is often preconditioned on the availability of DMT [28, 30] and that access to treatment can be conditional on the results of the screening. It is clear that some form of qualification for treatment access other than age is needed once preventive treatment is available [38] in order to avoid overmedicating the whole population and unsustainable costs, but no answers are given as the topic is addressed only very sparsely in the literature. In this context, the question emerges whether it is ethical to restrict the access to DMT based on the results of model-based assessment of AD risk, given that for some proportion of patients they might be false [30]. Subsequent considerations, that the model could also biased, unreliable or otherwise faulty are not being discussed in the AD prevention literature. We elaborate more on these topics in discussion.
Instead, the main themes that emerge around the topic of treatment and conditioning of treatment access is equity and distributive justice, understood mostly as equal access of individuals at risk of AD to general health services as opposed to being discriminated against by insurers [22, 54] and balance in the amount of stakeholders’ attention and resources dedicated to preclinical AD, versus dementia [12, 36], versus other healthcare needs [12, 52]. In addition to that, the burden incurred to the healthcare systems by addressing preclinical AD is of concern [4, 51].
Even once access to potential, future DMT is granted, an important uncertainty remains regarding the rationale for prolonged treatment in the preventive settings. There is a concern that possible side effects of preventive treatment [4, 36], coupled with intensive and potentially invasive monitoring might in some cases overweigh the benefits [21, 62]. Therefore, future patients need to be informed about the benefits and risks of treatment to be able to weight these factors in according to their own personal values and make an informed decision [28, 60].
The concern regarding the benefit of future preventive treatment is amplified by the uncertainty, whether a treatment benefit observed in clinical trials will represent the true effect in a real-world population of patients. This could happen if real-world patients are different, for example, more diverse than those recruited to the clinical trials based on stringent inclusion criteria [30]. A concern is raised also regarding whether the outcomes meaningful to patients will be adequately captured or at least informed by the clinical trials, which are typically limited in their time of follow-up [30, 52]. This short time horizon of clinical trials is being seen as the main hazard for the ability to make an accurate assessment of preventive treatment’s real-world outcomes and cost-effectiveness [22, 63]. Cost-effectiveness of both diagnostic tests and the preventive treatment is seen as a requirement for offering them to patients [54, 64] but the literature diverges when it comes to opinions whether future AD treatment will be cost-effective or not. Some papers present claims that future early treatment will be superior to current symptomatic treatments, and that it will offset costs of healthcare and institutionalization. In such scenario, there is even an ethical obligation to make this treatment available to patients [22, 40]. Opposite views dominate though due to a concern that the direct cost of innovative preventive treatment and of associated clinical monitoring will be large [22, 38], while offsets will occur in the social care, rather than healthcare system [30, 38]. Health-economic modeling can be used to resolve this dispute, however, there is a caveat in that modeling is highly complex and the results depend on modeling assumptions. Therefore, model inputs must be clearly defined and transparently communicated [38]. The literature does not provide answers yet as to what predictive power of a model used for preclinical testing would be desired and acceptable.
Broader social implications of using predictive models for AD prevention, and other social issues
The existing literature recognizes the need to facilitate development and adoption of effective AD strategies, given the major public health importance of AD. Public-private partnerships are often mentioned as an example of such strategies [12, 63]. A sense of urgency can be seen regarding the need to regulate access to AD risk assessment which is already available to some patients through direct-to-consumer testing [40, 63].
The future preventive approach to AD is expected to put a strain on the healthcare and social system, creating a demand for more intensive interaction between patient and doctor, assistance to people with preclinical AD to plan for and monitor emerging disabilities [22, 65], and to provide care arrangements for them [31, 38]. The existing literature recognizes the imminent tensions which might arise from this and calls for a priority setting process with public participation [28] and postulates that all patients in need have access to diagnosis and treatment, to prevent further health inequalities [22].
Further important ethical questions raised in the context of AD prevention using predictive modeling is how far medicine should go in terms of treating risk factors or risk status [12, 55], and to what extent it should become “clinical-actuarial rather than clinical-pathologic” [31, 58]. The rise of so-called “desktop medicine”, where patients learn about their health not based on their symptoms but test results introduce the need to appreciate technological challenges, e.g., to develop an optimal governance model for patient’s data, assure their privacy and accountability of those handling them [36, 44]. Not to be ignored are also high technical and infrastructural requirements for data gathering and managing, particularly for population-level AD screening [22, 38] and the need to adapt professional practices, social policies and legal infrastructure need to evolve to accommodate this paradigm shift in AD treatment [31].
Last but not least, it is worth noting that the topic of AD secondary prevention is discussed mostly from a perspective of high income countries, leaving out unanswered questions such as how these topics are being perceived outside the Global North [36, 40] and whether low and middle-income countries possess means and infrastructure to also benefit from early AD diagnosis, management, and treatment [38]. The expectation is that the transnational gap will only increase once DMT become available [38].
DISCUSSION
This review investigated the ethical and social considerations which arise in the secondary AD prevention setting where predictive models can be used particularly for assessment of the risk of AD clinical symptoms in cognitively unimpaired individuals or prediction of long-term AD outcomes with and without treatment. The themes drawn from the reviewed literature reflect current academic discussion of those aspects that bear ethical or broader societal relevance, i.e., can be understood as statements regarding ‘right’ and ‘wrong’, the ‘goodness’ of practice or phenomenon, or competing normative interests and values among relevant stakeholders.
Much, although not all [38], of the current literature is centered around the DMT being a necessary and sufficient condition for ethical risk assessment and disclosure. We did not identify articles discussing the benefit-to-risk of non-pharmacological interventions (e.g., cognition-based intervention, physical exercise) that may be effective in the early stages of AD [66], while better tolerated than pharmacological options. We argue that the discovery of DMT while resolving many critical issues related to AD prevention, creates others.
The first one is that the availability of a DMT will not automatically translate into accessibility, challenging the principle of equity and distributive justice. One can expect that such innovative treatment will be costly, at least during the first years after launch when it will be protected by a patent, and so will the battery of tests needed to select the target preclinical population. This means that a large proportion of patients who could benefit from preventive treatment, insured in middle- and lower-income countries, might not have access to it while in high income countries paying for AD preventive treatment will off-set other healthcare or public needs.
The other issue introduced by the discovery of DMT is that despite overall efficacy demonstrated in a clinical trial, some aspects of drug’s benefit-to-risk will remain unclear. This is because the long-term consequences of using this treatment will not be clear from the RCT alone, and because it will not be known whether all eligible patients will benefit from the treatment, and whether some patients will be harmed. This issue, though, rooted in the ethical principle of non-maleficence, can be mitigated with further post-marketing studies, monitoring long-term consequences of such treatment and further scientific progress in the identification of potential responders, possibly leading one day to a stratified or even personalized medicine approach to AD. On the other hand, a rush in introducing a DMT into clinical practice in the preventive setting might severely limit our ability to adequately and comparatively monitor the long-term progression of AD and the long-term benefit to risk of any further treatments developed thereafter.
Finally, a side effect of the attempt to alter the AD trajectory and postpone, or even prevent cognitive decline and disability, is its contribution to creating a new patient population of “worried well” from individuals who otherwise considered themselves healthy, to the medicalization of private life, and to transforming medicine into an “actuarial” science and practice. These changes are not trivial. Positive AD risk assessment can impact self-perceptions or self-identity. Similar effects can occur for relatives, family members and friends who discover information about their susceptibility to AD, or learn about the susceptibility of a relative, resulting in modification of familial, social, or caring roles. AD risk assessment may likewise result in discrimination comparable to that facing symptomatic AD dementia patients, family members, and carers. People at risk of AD (i.e., who may or may not develop AD dementia at some point in the future) may, for example, also be exposed to attitudes, practices or procedures which potentially devalue or discriminate against them (e.g., monitoring their ability to manage finances or to drive already before the symptoms occur, perhaps even as a part of a well-intended policy). This is while patients often fear loss of agency more than they fear death [43], perhaps because of the social stigma associated with AD overemphasizing the most advanced stages of AD, as opposed to providing support allowing people affected by AD to function in various domains in life as long as possible.
Another finding from this review is that although ethical issues in AD secondary prevention are discussed abundantly in the literature, specific issues related to modeling used to predict AD risk are not scrutinized. One instance of this is that the existing literature around disclosure practices seems to be deeply anchored in the paradigm of a single risk factor, primarily genetic, or to a lesser extent, biomarker-related. Assessment of personal risk estimated using advanced predictive methods, combining a number of patient characteristics as described above (e.g., demographics, genetics, brain images, blood biomarkers, and medical history) is scarce and incomplete. For example, the uncertainty around the prediction of AD is typically understood in the literature as the probability of making a false-positive diagnosis and therefore raises the problem of misclassification by a predictive algorithm. The reviewed literature is likewise missing any specific considerations regarding the clinically and socially acceptable levels of precision and reliability of the models which could be used in the AD secondary preventive setting and therefore, it is currently not possible to derive from the literature any indication about the qualities of a predictive algorithm that would justify its use in populations known to be at risk, and in the general population.
Also specific sources of uncertainty and biases leading to misclassification are not being discussed. Such biases can be purely technical (e.g., low granularity of data for prediction affecting the precision of prediction) but can also be rooted in social attitudes and practices, either pre-existing at the time when a predictive model is being developed or emerging during and through the use of this model [67]. As a hypothetical example, a person whose relative have AD might be more likely referred to a specialized memory clinic compared to a person without this risk factor (preexisting bias) resulting in data from memory clinics over representing this type of future patients (technical bias). In effect, a model developed on such data could produce more accurate predictions for this group of patients, compared to others (external generalizability). Such a model subsequently used in a clinical practice could then contribute to the underrepresented patients receiving suboptimal care or even to being discriminated against. The clinical use of such a complex, multivariate predictive model would pose more challenges. For example, the same level of risk can be derived from such a model for two patients based on completely different sets of characteristics, and therefore, be associated with a different degree of uncertainty. This feature might make the AD prediction based on such a model more demanding to communicate to both the patient and the treating physician. If machine learning was used to develop the model, which is increasingly the case, it would even be very difficult to trace back the reasons why a certain prediction was made. In additional to that, commercially developed models will likely be patented and not open for public scrutiny. Therefore, any potential harm caused by biased prediction would be difficult to discover, posing a risk that a faulty model would shape the clinical practice for an extended period of time and leading to a dispute who is to be held accountable for the fault of a self-learning, black-box predictive model [68]. In this new context, data governance needs to be reassessed, starting from fundamental issues such as informed consent (To what extent is it possible, given the complexity and unknown long-term consequences of using predictive models in routine care?) and data privacy (How to assure that patients will not be de-identified based on a unique set of characteristics used in the multivariate predictive model?), through ownership (If patients or clinics contributed data to develop a model, who owns the model?) and accountability (What business model would best strike balance between model developers rights and profits and public interest?), all the way to very specific consideration around data sharing for modeling purposes. The latter is a challenge because unlike in the case of descriptive analytics which can be generated internally within the institution of a data owner and shared externally, building a predictive model requires multiple iteration of access to data which can hardly be done without a physical access.
Furthermore, the existing literature on AD barely mentions a possibility that a predictive model can serve not only as an elective preventive procedure, but also as a basis for a populational surveillance system, e.g., when connected to an Electronic Medical Record (EMR) system. This is despite the growing interest in using EMR for public health surveillance and case detection [69, 70]. In AD setting, patients identified by one predictive algorithm with high sensitivity and low specificity could be called into a healthcare practice for an AD risk assessment, using a more specific algorithm, e.g., including biomarkers. Sparsely covered is the problematic of the large technical and infrastructural requirements of AD secondary prevention. Any extensive use of such advanced models predicting risk of future AD will have large logistic requirements for data collection, processing, and storage.
While these large themes are clearly underrepresented in the current literature on AD, a discussion around mathematical models used to predict future AD outcomes for the needs of health technology assessment is emerging. The most straightforward example of such a model is a health-economic model which will be needed to evaluate the cost-effectiveness of the preventive AD treatments, once they are developed. It is being recognized that results of such a model will depend to a large extent on the choice of the modelled outcomes and assumptions. Therefore, established criteria for such model’s trustworthiness are needed, so that it could be used for decision making. As part of this effort, a series of studies have been conducted in the ROADMAP project (Real World Outcomes across the AD Spectrum for Better Care) [71], focusing on the ethical and social implications of data sharing and repurposing, priority outcomes for different AD stakeholders and methodologies as well as input data used in the currently existing health-economic models in AD [72, 73]. Reporting standards have also emerged for both economic evaluations and predictive models (CHEERS and TRIPOD, respectively) [74, 75].
Future directions
This review uncovered several directions for future research.
The first one would be to supplement the current review conducted in the AD setting, with a review of literature on the developments in the field of predictive modeling, machine learning, and precision medicine, which—even if not specific to AD—could provide a perspective on specific challenges to be expected if predictive models are used in routine clinical care. Some lessons can also be learned from other specific fields were predictive modeling was applied to assess credit score, predict child abuse, criminal offence, among others [76]. Such a review could also further explore how the use of predictive models for preclinical risk assessment can affect access to preventive treatment. For example, whether risk stratification could lead to unfair exclusion of people who might desire to receive a preventive AD therapy, but be denied access because of not meeting a certain pre-defined risk threshold.
The second direction would be to examine the perspectives on secondary AD prevention from low- and middle-income countries, given that the reviewed literature discussed mostly the high-income countries perspective. Some of the differences which we expect to see would be in beliefs about the benefits of risk disclosure and in considerations and realities of limited access to current and future AD therapies.
Finally, another topic to explore is the possible policy consequences of a large-scale AD prevention. The literature suggests that focus on prevention would divert resources from care offered to symptomatic AD patients. It is, however, possible that standards of AD care improve, if large scale AD risk assessment creates an organized group of cognitively unimpaired people aware of their likely future with AD clinical symptoms and ready to engage in policy making.
Limitations
One potential limitation of this study is that it reflects the current status of the ethical discussion about the ethical aspects of using predictive modeling in AD secondary prevention. We found that this discussion does not yet follow the most recent medical developments in the AD field. Similarly, we did not identify articles discussing the benefit-to-risk of non-pharmacological interventions (e.g., cognition-based intervention, physical exercise) that may be effective in the early stages of AD, while better tolerated than pharmacological options [66].
Another potential limitation stems from the fact that the mapping of ethical themes relies to a large extent on qualitative interpretation of the reviewers. To mitigate the risk of self-nomination eight principles were used as guidelines to establish the ethical relevance of each theme. The eight principles are not intended as an ethical framework for predictive modeling, but rather were used as a reference point to further establish the ethical relevance of the themes identified in the reviewed studies beyond self-nomination by study authors.
Conclusions
Based on our understanding of the AD and therapeutic landscape in this indication, we believe that advanced predictive modeling might become an indispensable element of AD preclinical prevention. In such scenario, given the numerous ethical concerns associated with this approach, safeguards need to be implemented. Public health and medical institutions undertaking AD preventive programs are accountable to the general public and patient populations whose health and well-being are at stake. Risk-benefit assessments, model validation, and development of professional practices and norms are necessary to establish and deliver effective and publicly beneficial screening programs and treatment access plans. Evidence supporting the implementation of such programs should be shared with relevant patient populations to support well-informed autonomous decision-making regarding participation.
Footnotes
ACKNOWLEDGMENTS
This project has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement No 116020 (“ROADMAP”). This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation programme and EFPIA. The funder played no role in this study.
The authors would like to acknowledge the contributions of the ROADMAP consortium (
) and Dr. Alex McKeown from the University of Oxford for providing guidance on the development of the data extraction sheet.
Andrew Turner’s time is supported by the National Institute for Health Research Collaboration for Leadership in Applied Health Research and Care West (NIHR CLAHRC West), now recommissioned as NIHR Applied Research Collaboration West (NIHR ARC West). The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care.