The need for systems thinking to advance Alzheimer ’ s disease research

Despite extensive research efforts to mechanistically understand late-onset Alzheimer ’ s disease (LOAD) and other complex mental health disorders, curative treatments remain elusive. We emphasize the multiscale multi-causality inherent to LOAD, highlighting the interplay between interconnected pathophysiological processes and risk factors. Systems thinking methods, such as causal loop diagrams and systems dynamic models, offer powerful means to capture and study this complexity. Recent studies developed and validated a causal loop diagram and system dynamics model using multiple longitudinal data sets, enabling the simulation of personalized interventions on various modifiable risk factors in LOAD. The results indicate that targeting factors like sleep disturbance and depressive symptoms could be promising and yield synergistic benefits. Furthermore, personalized interventions showed significant potential, with top-ranked intervention strategies differing significantly across individuals. We argue that systems thinking approaches can open new prospects for multi-factorial precision medicine. In future research, systems thinking may also guide structured, model-driven data collection on the multiple interactions in LOAD ’ s complex multicausality, facilitating theory development

Late-onset Alzheimer's disease (LOAD) is a pervasive neurodegenerative disorder that adversely affects various cognitive functions, such as language, memory, planning, and visuospatial skills (Gaugler et al., 2022).As the most prevalent form of dementia, LOAD affects approximately 11 % of individuals aged 65 years and older in the United States (Gaugler et al., 2022), imposing a substantial and growing economic burden (Wong, 2020).Beyond financial implications, LOAD diminishes life expectancy, averaging around six years following diagnosis (Liang et al., 2021).The toll on quality of life is also profound (Wood et al., 2016), inflicting tremendous suffering on patients and their caregivers globally.
Given the urgency surrounding LOAD, treating or at least delaying LOAD onset and progression is of paramount importance.Unfortunately, current treatments for LOAD do not yet halt or reverse its course, with 98 % of phase-2 and phase-3 trials not finding significant effects (Kim et al., 2022).Even recent successful drugs have debated effectiveness and raised questions regarding safety and clinical relevance (Liu et al., 2021;Walsh et al., 2022).Furthermore, while large-scale multifactorial randomized controlled trials aimed at dementia prevention have shown promising results (Ngandu et al., 2015), their overall outcomes are mixed (Andrieu et al., 2017;Richard et al., 2019;van Charante et al., 2016).Combined, these research challenges underscore the complexity of LOAD and the need for further research.Therefore, in addition to the ongoing prevention trials (Kivipelto et al., 2020) and drug candidate studies (Cummings et al., 2023), novel paradigms are needed to advance our understanding of LOAD and develop new intervention strategies.

The multiscale multicausality of late-onset Alzheimer's disease
Although accumulation of the amyloid beta protein has been traditionally regarded as its hallmark process (Karran et al., 2011), LOAD is increasingly seen as a complex interplay between various pathophysiological mechanisms at multiple levels of biological organization (Fotuhi et al., 2009;Mullane and Williams, 2020;Sweeney et al., 2019;Uleman et al., 2020Uleman et al., , 2023b)).In this picture, protein accumulation is only one of many contributing mechanisms (McCleery et al., 2019;Pimplikar, 2009), with numerous other pathophysiological processes, including inflammation (Newcombe et al., 2018), vascular dysfunction (Sweeney et al., 2019), oxidative stress (Butterfield and Halliwell, 2019), and glucose metabolism (Butterfield and Halliwell, 2019).These processes are not limited to the cellular scale, as organ, individual, and societal scales are also involved (Fig. 1).
Beyond pathophysiological processes, a combination of genetic (particularly the ApoE-4 gene), environmental, and lifestyle factors also play a role in LOAD etiology.Modifiable risk factors like social isolation, physical inactivity, hypertension, and obesity account for approximately 40 % of the population's dementia risk (Livingston et al., 2020), or even higher in lower-income countries (Mukadam et al., 2019), emphasizing the potential for multifactorial prevention strategies (Kivipelto et al., 2018).By now, there is a wealth of research on LOAD risk factors and pathophysiological processes.However, our understanding of the intricate interplay among them is still unfolding (Fotuhi et al., 2009) as most LOAD studies have primarily focused on isolated pathophysiological mechanisms or considered only a small number of factors and relations.Although these studies are fundamental for delineating specific processes, the multicausality of LOAD is rarely studied as a whole (Kuljis, 2009), limiting our understanding of how these processes interact and collectively contribute to LOAD's etiology (Tang et al., 2019).
As a complementary paradigm, LOAD can be conceptualized as a multiscale and multicausal system, encompassing an array of pathological factors and interactions (McCleery et al., 2019;Pimplikar, 2009;Uleman et al., 2023b) to gain a more comprehensive understanding (Fotuhi et al., 2009;Kuljis, 2009).Systems thinking provides a methodology to achieve this by offering systematic tools for studying complex disorders from a complexity science perspective, holding great promise for unraveling the biopsychosocial etiology of mental health disorders and exploring intervention scenarios (Crielaard et al., 2022).In this paper, we discuss the recent contributions of these conceptual and computational methods to LOAD research and provide an outlook on their potential to advance the field.

Mental health disorders as complex systems
Systems thinking examines the interactions among various components ('holism') rather than focusing on specific components in isolation ('reductionism').It is important to clarify that systems thinking is not intended to favor holism over reductionism but rather to complement reductionist studies and create connections between them, fostering new insights (Fang and Casadevall, 2011).Indeed, studying specific components and pathways is a necessary step in systems thinking, as it provides the pieces of the systems map and yields data on how these pieces behave.However, to analyze complex systems appropriately, it is best to consider the interactions between pieces as part of analyzing their dynamics or making predictions regarding intervention effects.
To illustrate, the common amyloid cascade hypothesis (Karran et al., 2011) posits that amyloid beta aggregation leads to tau aggregation and neuronal dysfunction with consequent dementia (Fig. 2A).Many controlled experiments have been conducted to assess these pathways in isolation (Karran et al., 2011).However, to understand the broader system, including the various other putative upstream and downstream processes and risk factors, the link should be ultimately placed back in its multicausal context.A systems thinking approach would incorporate these other components into the picture, such as neuroinflammation (Tejera and Heneka, 2016) and oxidative stress (Haque et al., 2019), focusing on the interconnections and feedback loops (Uleman et al., 2020) (Fig. 2B).Indeed, LOAD shows features of complex systems, consisting of many components forming numerous connections and feedback loops (Ladyman et al., 2013), which can provide stability under perturbations (e.g., the adaptive cellular stress response (Mattson and Arumugam, 2018)) but may also trigger vicious cycles, such as sleep disturbance and depressive symptoms reinforcing one another, resulting in accelerated cognitive decline (Fang et al., 2019).
Like LOAD, many mental health disorders have limited treatment options despite extensive research efforts.These disorders typically involve numerous interacting mechanisms across multiple scales in space and time and are poorly explained by isolated mechanistic descriptions.Moreover, these interacting processes often behave nonlinearly (e.g., sudden changes in symptomatology around tipping points (van de Leemput et al., 2014)) and involve complex interactions with other disorders (e.g., multimorbidity with circadian rhythm disruptions (Alachkar et al., 2022)).Current tools and paradigms may not adequately capture the multiscale and multicausal nature of these complex disorders.Consequently, an increasing number of researchers recommend conceptualizing such disorders (Cramer et al., 2016;Hayes and Andrews, 2020;Kenzie et al., 2017;van der Wal et al., 2021), including LOAD (Cohen et al., 2022;Pomorska and Ockene, 2017;Rollo et al., 2016;Tang et al., 2019) as emergent properties of complex dynamical systems that cannot be reduced to their constituent parts and that may be more effectively addressed through systems thinking methods.

Systems thinking methods: causal loop diagrams and system dynamics
Systems thinking can be seen as a set of methods for understanding, reasoning about, and influencing complex systems.It involves a mix of qualitative and quantitative approaches to maximize the knowledge used when studying these systems since any single data source (e.g., expert knowledge and empirical data) often lacks critical information on the many different mechanisms involved.Although many definitions exist (Arnold and Wade, 2015), a systems thinking approach typically aims to understand the structure (components, connections, and feedback loops) underlying a complex system and simulate its behavior over time and the effect of interventions.To achieve this, systems thinking methods involve the development of one or more models.Although many types of models are used in science, such as animal models (e.g., APP23 mice for studying amyloid beta deposition (Yokoyama et al., 2022)) and physical models (e.g., the metal model of DNA (Schaffner, 1969)), systems thinking focuses on conceptual (i.e., a mental or visual representation of concepts) and computational models (i.e., a mathematical representation implemented in a computer).By combining conceptual and computational models, systems thinking facilitates the analysis of quantitative and qualitative data to delineate the complexity of LOAD pathophysiology and facilitate mechanistic insights and hypothesis development (Hampel et al., 2018a).
Conceptual modeling is often the firstand sometimes onlystep in a systems thinking approach and generally involves participatory methods, such as group model building (Crielaard et al., 2022;Uleman et al., 2020;Vennix, 1999).These methods require a collaborative effort of multiple scientific disciplines, making systems thinking approaches inherently interdisciplinary.To conceptually describe which system parts are important in a complex problem, it is common to draw a visual representation of the system that produces the problem, i.e., a causal loop diagram (CLD).CLDs are conceptual system models consisting of variables (representing, e.g., risk factors and neuropathological mechanisms) and assumed causal links between them, which can form feedback loops (Bala et al., 2017).These diagrams can provide valuable insights into a disorder's etiology and offer guidance for system interventions, such as disrupting detrimental feedback loops (Littlejohns et al., 2018).
Although such visual representations can provide insights, they cannot be used to reliably predict the outcome of system changes (Crielaard et al., 2023;Crielaard et al., 2022).For example, while a CLD may indicate that neuroinflammation is relevant in LOAD (Fig. 2B), this graphical representation does not quantify the potential reduction in cognitive decline after the inflammation is mended.To address such questions, a computational model is needed that translates the conceptual knowledge embedded in the CLD into simulations describing different intervention scenarios.This is critical because predicting the long-term behavior of a complex system based solely on its structure, particularly when involving feedback and nonlinearity, is inherently challenging for humans (Meadows, 2008).Therefore, a quantitative assessment of the system's dynamic behavior requires the conversion of a qualitative CLD into a computational model (Crielaard et al., 2022).
Various computational models have been developed to study LOAD at specific scales.For example, one model captured interactions among brain cells and protein aggregation, simulating the potential effects of pharmacological interventions, such as TNF-alpha inhibitors and antiamyloid beta therapy (Hao and Friedman, 2016).Subsequent research incorporated patient-specific disease trajectories calibrated on clinical data, predicting the efficacies of donanemab and aducanemab in slowing cognitive decline after amyloid beta clearance (Hao et al., 2022).Other computational approaches to LOAD modeled cognitive impairment with pathologic biomarker cascades over time (Petrella et al., 2019;Zheng et al., 2022) while also evaluating treatment scenarios (Petrella et al., 2019).These examples, which focus on the cellular and organ scales, highlight the significant contribution computational models can make in guiding LOAD research.
However, to integrate cellular interactions and biomarker dynamics with higher-level mechanisms, like lifestyle factors, computational systems models can be developed.The preferred choice for this purpose is often a system dynamics model (SDM) (Crielaard et al., 2022), although alternative complexity science methods, such as agent-based models, are also possible.By quantifying the components and connections from a CLD, SDMs can assess, for instance, how risk factor exposures influence pathological processes and subsequently result in cognitive decline.SDMs can also simulate the changes and interactions among different mechanisms over time and assess the effectiveness of interventions (Crielaard et al., 2022).

The need for data-driven system dynamics
While the call for systems thinking in psychiatry dates back at least forty years (Grinker, 1975;Marmor, 1983), the broader integration of systems thinking methods only gained momentum in health research over the past two decades (Darabi and Hosseinichimeh, 2020).Acknowledging a problem's complexity without a corresponding modification in the methodological approach represents a 'weak' complexity commitment, which can be strengthened by adopting systems methods like CLDs and SDMs (Hasselman, 2023;Hasselman, 2022).Despite the growing adoption of these methods, only a few studies specifically focus on ('individual level') mechanistic disease modeling, presenting a significant opportunity (Darabi and Hosseinichimeh, 2020).
Such mechanistic disease models, as demonstrated by pioneering applications on concussion (Kenzie et al., 2018;Wakeland and Kenzie, 2019) and depression (Hosseinichimeh et al., 2018;Wittenborn et al., 2016;Wittenborn and Hosseinichimeh, 2022), are often constrained by limited data availability, hindering the conversion of CLDs into SDMs (Crielaard et al., 2022;Sterman, 2018).For instance, a detailed CLD was developed for depression, whereas a more focused data-driven SDM was implemented to address stressors and rumination, limiting its applicability to adolescents (Hosseinichimeh et al., 2018;Wittenborn and Hosseinichimeh, 2022).Conversely, while a comprehensive CLD for concussion was successfully converted into an SDM, it lacked calibration due to insufficient longitudinal data, making it speculative (Kenzie et al., 2018;Wakeland and Kenzie, 2019).
These approaches offer insights into specific aspects of these complex disorders.For example, the SDM for depression identified distinct depression profiles in adolescents (Hosseinichimeh et al., 2018)   2022).The SDM for concussion highlighted the importance of physiological processes through sensitivity tests, underscoring the need for multi-dimensional data that include relevant biomarkers (Wakeland and Kenzie, 2019).However, these approaches do not suffice for simulating interventions across a range of risk factors and pathophysiological processes in diverse real-world individuals.To achieve this, a comprehensive mechanistic disease SDM should be developed using large multi-dimensional data sets.Inferences from less comprehensive models may introduce biases by omitting relevant mechanisms and potential confounders.
Recent advances present new opportunities for LOAD and dementia research, where systems thinking methods were only recently introduced to unravel disease mechanisms (Seifert et al., 2022;Uleman et al., 2023bUleman et al., , 2023aUleman et al., , 2020)).Efforts have been made in the psychological literature to maximize expert knowledge and scientific literature in addition to empirical data when developing biopsychosocial CLDs and SDMs (Crielaard et al., 2022).Moreover, large amounts of multi-dimensional data have also become increasingly available through platforms like GAAIN (GAAIN, 2020) (e.g., cognitive-behavioral tests, lifestyle factors, brain imaging, and biomarkers).Therefore, LOAD studies have started addressing the abovementioned challenges by converting a comprehensive CLD into a data-driven system dynamics model, which can serve as a case study for the broader context of mental health research.

A case study in systems modeling of Alzheimer's disease
Over the past few years, our research group employed systems thinking methods to synthesize diverse research findings, enhance our understanding of LOAD's multicausal etiology, and identify promising targets for multifactorial LOAD prevention (Uleman et al., 2023b(Uleman et al., , 2023a(Uleman et al., , 2020)).We first developed a CLD through a group model building process with 15 domain experts from various scientific disciplines, including neurology, gerontology, neurobiology, psychology, and neurochemistry.This collaborative effort and an extensive literature review culminated in a comprehensive CLD for LOAD (Uleman et al., 2020).The CLD encompasses 38 causal variables in LOAD and 150 connections between them.For example, highly connected variables in the diagram included social relationships, cerebral endothelial dysfunction, physical activity, circadian misalignment, and systemic inflammation.Crucially, the CLD revealed several reinforcing feedback mechanisms that potentially play a critical role in LOAD onset.Similar to the previous CLDs developed for concussion (Kenzie et al., 2018) and depression (Wittenborn et al., 2016), some of these loops span multiple spatial scales, such as sleep quality (individual), glymphatic system dysfunction (organ), tau burden (cell), neuronal dysfunction (organ), cognitive decline (individual), and back to sleep quality.These cross-scale loops could be overlooked when approaching the disorder from a single scientific discipline alone (Kenzie et al., 2018), emphasizing the need for interdisciplinary approaches to address complex mental health disorders effectively.
Following the development of the CLD, we sought to quantify the system and identify intervention targets.This was achieved by converting the CLD into an SDM using two observational longitudinal LOAD data sets: the Alzheimer's Disease Neuroimaging Initiative (ADNI) from the United States (N = 1762) and the Hellenic Longitudinal Investigation of Aging and Diet (HELIAD) from Greece (N = 1924) (Uleman et al., 2023b).We then simulated 5-year cognitive decline trajectories and, as a first model validation step, ranked the effects of potential interventions on 15 modifiable risk factors.This ranking compared well to meta-analyses of randomized clinical trials and observational studies, suggesting that the top-ranked factors, sleep disturbance and depressive symptoms, may be opportune intervention targets (Uleman et al., 2023b).This effect was partly mediated by the abovementioned cross-scale loop involving sleep quality and tau burden (Uleman et al., 2023b).Ongoing prevention trials that address sleep and mental health, such as PREVENTION (McEwen et al., 2021) and FINGER-NL (Zwan et al., 2021), are expected to provide further insight into the validity of these predictions.

Multifactorial precision medicine in Alzheimer's disease
While integrating interventions for sleep and depressive symptoms into other prevention studies may be promising, a precision medicine approach may be required to best deliver interventions to individuals (Isaacson et al., 2019).Indeed, population-aggregated models might not adequately capture individual mental health trajectories, emphasizing the necessity for a person-specific approach (Hasselman, 2023).Precision medicine is advocated extensively in LOAD research, notably by the Alzheimer Precision Medicine Initiative (Hampel et al., 2019).Consequently, precision medicine approaches are on the rise, exemplified by advancements in pharmacology (Hampel et al., 2018b), including targeting neuroinflammatory pathways (Hampel et al., 2020).
Recent preliminary proof-of-concept clinical trials demonstrate the potential of extensive personalized prevention programs in averting cognitive decline (Bredesen et al., 2023;Isaacson et al., 2019;Yaffe et al., 2023).However, such trials are resource-intensive (e.g., time, cost, and labor) and, therefore, necessitate protocol refinements and simplifications (Bredesen et al., 2023), especially considering the limited feasibility of adherence to numerous lifestyle changes (Isaacson et al., 2019).Thus, identifying an optimized set of individualized interventions may be crucial to bridge the gap between efficacy and practicality.While acknowledging that real-world plans can never be optimal, such personalized interventions are expected to surpass the efficacy of population-average interventions.Personalized SDMs could thus come to play a role in clinical decision-making by simulating personalized interventions (Rogers et al., 2018).
To explore this idea, we extended the SDM for LOAD to simulate personalized and multifactorial interventions using both the ADNI data and an additional data set: the Australian Imaging, Biomarker & Lifestyle Flagship Study of Aging (AIBL) (N = 1903).This extension incorporated interaction terms between different model variables and estimated individual-based parameters, leading to several key insights (Uleman et al., 2023a).First, the multifactorial intervention simulations suggested that concurrently targeting sleep disturbance and depressive symptoms may yield synergistic benefits, surpassing the combined effects of intervening on either alone.Second, the most promising risk factors varied widely across persons (Fig. 3), highlighting the need for a precision medicine approach in LOAD prevention to target these specific factors.Third, we assessed the potential benefit of personalized interventions compared to group-level ('one-size-fits-all') interventions.Only in 13 % of individuals the highest-ranked intervention was equivalent to the group-level's.The remaining 87 % had a statistically significant benefit with a median of 3.5 ADAS-cog-13 points [95 % confidence interval of the median CI: 3.1, 3.8] ( Uleman et al., 2023a).In the top 8 % of individuals for whom a personalized approach was most advantageous, this benefit was much higher: 14.6 [12.0, 16.1] ( Uleman et al., 2023a).Although these preliminary findings require further calibration and validation through external data sets and N-of-1 trials (i.e., single-patient studies focused on individual responses to different interventions), they hint at the potential of SDMs as tools for precision medicine.Consequently, SDM-guided designs may yield greater efficacies than those observed in current multidomain prevention trials.As a subsequent step, the model could also extend beyond a prevention context, simulating pharmacological interventions on the various embedded pathophysiological processes, including amyloid beta and tau immunotherapies.

From knowledge synthesis to system-level theories
In closing, systems thinking holds the potential to advance LOAD research by synthesizing and appraising diverse scientific findings, Fig. 3. Top 3 single-and double-factor simulated intervention effects in two individuals.This figure is based on previously published simulation data of a 94year-old male and a 55-year-old female (Uleman et al., 2023a).The effect sizes are operationalized as the difference in the cognitive subscale of the Alzheimer's Disease Assessment Scale comprised of 13 modalities (ADAS-cog-13) after 12 years with and without an intervention on one or two modifiable risk factors.In the case of a double-factor intervention, the intervention strength of a single-factor intervention (1 standard deviation) was spread out over both factors (i.e., ½ standard deviation each).The bars are medians with 75 % percentile intervals of the median.guiding the way toward (personalized) interventions and new empirical studies.Represented as an hourglass (Fig. 4), the approach synthesizes extensive quantitative data and scientific literature on complex mental health disorders into a comprehensive CLD by domain experts.The CLD, once quantified, serves as the foundation for an SDM, generating hypotheses through simulated (personalized) interventions.Reversing the hourglass, the SDM outcomes can guide clinical, translational, and basic studies, facilitating the collection of the multi-dimensional data needed for attaining novel mechanistic insights (Roach et al., 2022).In addition to (external) validation, these studies can focus on reducing model uncertainty to incrementally improve the model's precision (Uleman et al., 2023a).By closing the loop between existing knowledge, computational modeling, and empirical follow-up research, systems thinking enables an iterative process of developing system-level theories for LOAD and other complex mental health disorders.

Conclusion
This paper makes a case for broadly adopting systems thinking methods in studying LOAD and other complex mental health disorders.Our argument highlights the potential of modeling LOAD as a multicausal system and using an SDM to simulate personalized interventions, resulting in insights that can guide further research.Although further optimization and validation of these models are needed, the first proofs of principle are encouraging.For example, we illustrated the utility of system dynamics in simulating personalized, multi-factor interventions, encouraging similar approaches for other complex disorders.This approach holds the potential for more effective precision medicine in clinical practice and offers a holistic perspective on this devastating disorder.We hope this recent computational research inspires a shift towards systems thinking, fostering new hypotheses, interdisciplinary collaborations, and an optimistic outlook on future research into LOAD and other complex mental health disorders.

Fig. 1 .
Fig. 1.The multiscale complexity of late-onset Alzheimer's disease with example mechanisms at different spatial scales.

Fig. 2 .
Fig. 2. Reductionist thinking compared to systems thinking at the level of the brain.This example is based on a causal loop diagram for late-onset Alzheimer's disease that also incorporates other relevant biopsychosocial factors (Uleman et al., 2020).