Data for this study were obtained from the HHS (Figure 1) [16]. The initial survey was conducted between 2000 and 2002, targeting 40- to 60-year-old employees of the City of Helsinki, Finland (phase 1, N=8960, response rate 67%). Follow-up surveys were administered to all phase 1 participants in 2007 (phase 2, response rate 83%), 2012 (phase 3, response rate 79%), 2017 (phase 4, response rate 82%), and 2022 (phase 5, response rate 75%). The final analytic sample comprised 8960 individuals.

Outliers were treated as missing values. We generated descriptive contingency tables to illustrate missing data and the sizes of each variable group across phases 1-5. Due to the infrequency of missing covariate data and the capability of Bayesian networks to handle such values, we did not perform any imputations [29]. Moreover, the full dataset could be used in the model, as each node’s conditional probability table (CPT) was estimated based on the available cases for that specific variable, without requiring complete data across all variables. All categorical variables were treated as discrete random variables, with domains that were mutually exclusive and exhaustive.

Initially, we established the model structure using expert knowledge and automated techniques. Temporal variables (age group, consumption of fruit and vegetables, smoking, alcohol consumption, LTPA, BMI, and insomnia symptoms) were hardwired between time slices to represent development paths. We then used a score-based quotient normalized maximum likelihood (qNML) criterion algorithm to identify additional relationships within the same time slice (horizontal effects) between nodes. qNML is an information-theoretic model selection criterion similar to the commonly used Bayesian information criterion (BIC) [30]. It identifies the most informative parent sets for each variable, while penalizing model complexity. qNML has been found to favor models that, although parsimonious, tend to achieve higher predictive performance than models selected by BIC for moderate and large sample sizes [31]. In the qNML algorithm, the nodes were ordered as follows: age group, gender, education, consumption of fruit and vegetables, smoking, alcohol consumption, LTPA, insomnia symptoms, BMI, hypertension, high cholesterol, diabetes, psychiatric disorders, current pain status, concentration, memory, and learning. The decomposable qNML algorithm facilitates separate structural learning for each node and its potential parents. To manage complexity, we limited each node to a maximum of 3 parents. Variables were organized so that causative variables preceded effects, but constructing a causal network was not our intent [32]. We assessed combinations of up to 3 parents for each node, selecting the configuration with the highest qNML score. Next, we derived CPTs for each node using our complete dataset. To address the zero-cell problem in probability estimation, we applied the Bayesian smoothing algorithm with a Dirichlet prior (1/2,...,1/2) to each CPT.

The use of a Bayesian network offers significant advantages, primarily the ability to perform inference based on specified evidence and events. Age is the initial parameter in this dynamic model. For the initial time slice, CPTs represented prior probabilities from our dataset. This inference mechanism supports a scoring calculator, where users input specific data into the network. This allows the calculation of event probabilities, such as the likelihood of memory decline or probabilities related to the longitudinal development of variables. Here, “evidence” referred to user-provided data used to infer the HHS score based on a pretrained DBN.

The expectation propagation algorithm, a deterministic method, was used for inference, yielding approximate results in discrete cases [33]. In inference, information contained in the evidence provided by the user is propagated through the network one variable at a time, updating the probability distribution of each variable by conditioning on the evidence. We calculated the full posterior distribution for each variable within each time slice (t) to maintain uncertainty and probabilistic relationships among variables. Results were presented with Bayesian credible intervals, which inherently account for uncertainty and ensure robust estimation of relationships, while preserving probabilistic dependencies over time. When transitioning from one time slice to the next, we used the Maximum A Posteriori (MAP) estimates of the variables from the previous time slice (t – 1) as inputs for calculating the current time slice (t) to reduce computational complexity and prevent the overpropagation of uncertainty. Cross-correlations and autocorrelations among output variables were also calculated from our dataset, based on demographic and risk factors.

We validated our model using 5-fold cross-validation and evaluated classification performance by calculating the area under the receiver operating characteristic curve (AUROC) [34].

We introduced a dynamic decision heatmap as a communication tool, for instance, in patient consultations. This heatmap displayed the current state of a patient’s lifestyle and presented a decision matrix for each modifiable risk factor. It illustrated the effects of potential changes, projected 5 years into the future, showing risk changes derived from the DBN. When selecting a target intervention, the heatmap initiated precalculation in the DBN using the chosen intervention as evidence, generating new estimates for SCD to facilitate informed decision-making.

This study was approved and received a research permit from the health authorities of the City of Helsinki. The study plan was approved by the Ethical Committee of the Faculty of Medicine, University of Helsinki, Finland. Permission for the secondary use of health and social data was granted by the Finnish Social and Health Data Permit Authority, Findata. Participants provided informed consent for scientific research. All data were handled in compliance with data protection regulations.

The study population was based on survey data from the HHS (2000-2022). The baseline (phase 1) study population consisted of 40- to 60-year-old employees of the City of Helsinki, Finland (Table 1). Multimedia Appendix 2 shows the response rates by variable and study phase. Of all participants, 1842 (31%) reported a decline in memory, 2818 (47.4%) in learning abilities, and 1828 (30.7%) in concentration in 2022 (Table 2). Nondaily consumption of fruit and vegetables was reported by 892 (15.4%) to 2120 (23.9%) of the participants, while 3346 (48.2%) participants in phase 2 and 3094 (53%) participants in phase 4 were current or former smokers. Additionally, 575 (10.1%) to 1339 (19.0%) of the participants engaged in high or high-risk alcohol consumption. Physical inactivity was reported by 1650 (22.7%) to 1776 (26.5%) of the participants. At phase 5, 2183 (37.2%) had overweight, and 1312 (22.4%) had obesity. Finally, at phase 5, insomnia symptoms occurring >14 nights per month were reported by 1431 (27.2%) of the participants (Table 3). Furthermore, in phase 5 a total of 2529 (44.3%) of the participants reported pain, 2999 (59.8%) reported hypertension, 2608 (56.2%) reported high cholesterol levels, 837 (20.6%) reported diabetes, and 827 (20.4%) reported mental disorders (Table 4).

The structure of the DBN model, trained using qNML, is illustrated in Figure 2. The most informative parent nodes for the nodes representing memory, learning, and concentration were gender, alcohol consumption, and LTPA. These same parent nodes were selected for the nodes representing pain, hypertension, high cholesterol, diabetes, and mental disorders.

Analysis of the CPTs presented in Multimedia Appendix 3 indicated that LTPA consistently reduces the probability of SCD. Specifically, the average relative risk reduction of experiencing declined memory when transitioning from inactive to active physical activity states was 26.9% for men and 24.8% for women across different levels of alcohol consumption. Additionally, the relationship between alcohol consumption and SCD followed a U-shaped curve, with both nonconsumption and high-risk consumption exerting greater effects than moderate or elevated consumption. Manual testing of the network, corroborated by data from our sampled participants, supported these findings.

The user interface of the HHS score calculator is presented in Figure 3. The inputs for the calculator were set as evidence for Bayesian inference. The probabilities for each output, including developmental paths and risks, were read from the inferred model.

We selected an individual example from our sample to demonstrate the use of the HHS score calculator (Figure 3). This individual was a 55-year-old woman with secondary education who did not eat fruit and vegetables daily, currently smoked, did not consume alcohol, was physically inactive, with overweight, and had high insomnia symptoms. In a consultation with a health care professional, the decision heatmap (Figure 4) was used to communicate with the individual and propose personalized lifestyle modifications. For instance, a strategy focused on increasing LTPA over the next 5 years appears to be particularly effective. As shown in Figure 3, the outcomes indicated that, for example, the likelihood of subjective memory decline decreased from 40.2% to 33.9%, with an absolute likelihood reduction of 6.3 percentage points and an odds ratio (OR) of 0.76 (95% Bayesian credible interval 0.59-0.99). Considering that continuing the current lifestyle was estimated to lead to a 49.2% likelihood of memory decline, an intervention resulting in a reduction of 15.3 percentage points was estimated to decrease this likelihood to 33.9% (Figure 4).

We further found statistically significant cross-correlations between a reduction in SCD and various commonly reported health conditions, including hypertension (P<.001), high serum cholesterol (P<.001), and diabetes (P<.001), based on a sample at the initial time step, t (Multimedia Appendix 4). Mental disorders and pain showed minor contributions to the reduction in SCD. Following the intervention, the effects intensified at time step (t + 1), and subsequent autocorrelation analysis indicated stabilization from (t + 1) to (t + 3).

The model’s predictive power measured by the area under the curve (AUC) metric was moderate; the cross-validated AUC scores were in the range of 0.52-0.65, as shown in Multimedia Appendix 5. This indicated that the model provides meaningful estimates of average risk differences across lifestyle profiles.

We demonstrated how an AI approach with DBNs can be used to build a risk score for forecasting SCD, effectively handling complex interactions between different factors, and providing clear, interpretable outcomes. We found that LTPA appears to have the most consistent associations with subsequent SCD. In addition, we discovered that SCD and disease risks, such as diabetes, hypertension, and high serum cholesterol, are highly correlated.

The nature of SCD is multifactorial and may arise from normal aging (as older adults are more likely to monitor their memory and raise concerns), be reversible, or progressively lead to dementia [8]. SCD is prevalent in aging populations, and prevention of dementia is a global priority for the World Health Organization (WHO) [35,36]. Additionally, a lifestyle intervention program has been found to have beneficial effects for preventing cognitive decline in at-risk older adults [37]. This highlights the importance of the preventive HHS score, which addresses multifactorial issues through advanced AI-based DBNs to analyze complex data and identify emerging risk groups already in midlife.

Our study shows the complex associations between SCD and alcohol use, smoking, LTPA, consumption of fruit and vegetables, BMI, and insomnia. We found that changing these potentially modifiable factors could decrease the overall likelihood of SCD. Notably, LTPA emerged as the most significant predictor in our model, decreasing the risk of SCD. Numerous interventions have aimed to enhance cognitive functioning. For example, a systematic review by Barha et al [38] showed that, especially, aerobic exercise has a positive impact on cognition (standardized mean difference 0.85, SD higher compared to controls) [38].This effect size aligns with previous cohort studies, where high physical activity in mid-to-late life has been associated with up to a 38% lower risk of SCD [39] and physical inactivity has been linked to increased odds of SCD [40]. This finding aligns with the well-documented benefits of LTPA on cognitive function [41,42]. Other studies also corroborate that physical inactivity is a significant risk factor for SCD [43,44]. Many studies demonstrate the preventive importance of physical activity in achieving healthy aging and preventing frailty [45]. Increased physical activity can improve cerebral blood flow, enhance synaptic plasticity, facilitate cellular cleanup processes, decrease beta-amyloid levels, improve mental health, and reduce stress [46]. The novelty of this study lies in a new tool that assesses how changes in risk factors impact SCD during follow-up and how increased physical activity, for example, can reduce its likelihood.

The observed correlation between SCD and other disease risks may indicate shared underlying mechanisms or common risk factors. This suggests that interventions targeting cardiovascular risk factors could concurrently improve SCD [47]. Cognitive decline is also influenced by genetic factors, such as the ApoE4 genotype, as well as health conditions, which may interact with modifiable lifestyle factors [7]. Future studies should explore these interactions to better understand their role in the risk for SCD.

Given the shared risk factors, insights into SCD can provide valuable indications about potential progression to dementia. However, since our findings are based on observational data, the projected associations illustrated by the decision heatmap and the HHS score calculator should not be interpreted as direct evidence of intervention efficacy without confirmatory intervention studies. Although the longitudinal design, in which risk factors were assessed prior to the onset of SCD, supports the possibility of a causal relationship, it does not confirm that modifying these risk factors would influence the outcome. For that, randomized controlled trials (RCTs) are needed. However, the follow-up time in RCTs is often short, and the long-terms effects of interventions are difficult to capture, particularly for health outcomes that develop slowly (eg, dementia). Furthermore, it is possible to produce findings like RCTs even with observational data [48]. Prior to broader implementation, it is crucial to conduct an intervention study focusing on the identified risk factors to determine whether risk factor modifications can, indeed, lower the risk for SCD or dementia in line with the results suggested by the algorithm and what the adherence is to different lifestyle interventions in the long run.

The HHS score partially overlaps with existing dementia risk scores in terms of included lifestyle-related predictors. For example, the CAIDE (Cardiovascular Risk Factors, Aging, and Dementia) score incorporates the BMI and physical inactivity but not smoking, alcohol, diet, or sleep-related factors [49], while the LIBRA (Lifestyle for Brain Health) score covers most lifestyle components—smoking, alcohol, diet, and physical activity—but does not include insomnia [50]. Unlike these additive risk models estimating long-term dementia risk, the HHS score uses a DBN to predict SCD based on longitudinal data and interdependencies between variables. This approach allows for more individualized, time-sensitive predictions. A recent review emphasized the need for more dynamic, personalized tools for cognitive risk estimation, particularly those accounting for time-varying exposures [50].

In clinical settings, our score calculator could assist in assessing the probabilities of SCD. Additionally, our decision heatmap can support health care professionals when discussing how lifestyle changes could impact health outcomes with their patients. Our methodology integrates decision theory—combining elements of probability theory and utility theory—to evaluate the effectiveness of preventive measures [51]. For example, our DBN identified a U-shaped curve in alcohol consumption, similar to a large meta-analysis [52]. This finding exemplifies how latent effects can influence risk assessments and should be interpreted with respect to probabilities. This integrated approach ensures that recommendations are personalized and rooted in a comprehensive assessment of each patient’s unique health profile and needs. To support real-world clinical use, potential implementation barriers, such as integration with electronic health records, time constraints during appointments, and provider training, should also be considered. However, given the moderate predictive power (AUC 0.52-0.65), the HHS score should currently be considered a supportive tool rather than a definitive diagnostic instrument. Further external validation and refinement are needed before clinical implementation.

Our study has a few limitations. First, our attrition analysis [17] indicated that individuals with SCD are more likely to drop out in subsequent phases. Second, our results cannot be directly generalized to the general population, as the cohort was female dominated and comprised individuals who were employees of the City of Helsinki at inclusion. Although this homogeneity reduces confounding, it may limit external validity. The HHS score may perform differently in populations with more diverse socioeconomic backgrounds, in the general population, among people with different occupations and health care access, or in male-dominated populations. Importantly, gender was identified as an informative parent node for several other variables in the model. This means that due to the underrepresentation of men in the cohort, the CPTs for male-specific paths in the network were estimated from fewer observed values, potentially affecting the model’s performance and stability in male-dominated or gender-balanced populations. Future studies should validate the model in more heterogeneous samples.

Despite these limitations, our study also has significant strengths. First, our >20-year longitudinal cohort used identical questionnaires across all 5 phases, achieving moderate-to-excellent response rates for a broad range of sociodemographic and health-related variables. Nonresponse analyses confirmed that the data accurately represented the target population [16]. Second, we used an AI approach to better account for the complex dynamics and interactions among our variables. A DBN is an ideal model for our calculator because of its ability to process various types of data. It can effectively manage cases where some nodes are in an undefined state, labeled as “NA,” without compromising its functionality. This model simplifies the complexity typically encountered in networks by avoiding connections between every pair of variables. Instead, it links variables only across consecutive time slices. Unlike deep learning, which often requires large datasets and can suffer from reduced interpretability, DBNs are well suited for modeling structured relationships in smaller datasets. Additionally, DBNs allow for the simultaneous analysis of dynamic interactions among multiple variables over time, rather than assuming linear relationships as in traditional regression models. Additionally, the DBN model is characterized by its transparency, allowing users to inspect and understand internal processes. Moreover, it can handle incomplete datasets without requiring data imputation to fill missing values. However, DBNs are not without limitations. They can be sensitive to data quality, particularly when dealing with small sample sizes or noisy data, and they may be prone to overfitting if not properly regularized. Third, a key strength of the study is that we were able to conduct a first validation of the SCD against objective register data on dementia (medication data and hospitalizations). Our validation demonstrated a strong association between SCD and clinically verified dementia diagnoses, with individuals reporting memory decline being over 3 times more likely to have dementia in 2017 (age 57-77 years), and this risk increased to more than 5 times by 2022 (age 62-82 years). Additionally, declines in learning and concentration showed similar trends, reinforcing the role of SCD as an early indicator of later dementia risk. The receiver operating characteristic curve (ROC) analysis further confirmed its predictive validity, with an AUC of 0.78 in 2017 and 0.75 in 2022. Our measurement of SCD had good internal consistency, and consistency remained between phase 4 (2017) and phase 5 (2022). The consistent trends observed across phases suggest that the measure reliably captures the associations between SCD and many health-related factors and dimensions of health functioning. These findings confirm that SCD is a valid and reliable outcome measure, with significant prognostic value for dementia (Multimedia Appendix 1). This helps confirm our results and supports the future use of survey-based SCD when examining cognitive functioning and its risk factors.

At the individual level, the HHS score can provide concrete examples and motivation for understanding one’s future potential developmental patterns of risk factors and the contributions of their changes to later SCD. It could be used to detect the risk of adverse development early, especially when lifestyle information is included in the risk assessment by the patient or a health professional. Early detection and interventions already in midlife can improve the quality of life and add healthier years to life. From a broader perspective, these findings emphasize the need for AI-driven tools in public health and health care policy. Predictive models, such as the HHS score, could be integrated into preventive health care initiatives, supporting policymakers in designing evidence-based interventions to reduce the burden of cognitive decline. By using personalized health assessments, health care systems can proactively address modifiable risk factors and improve long-term cognitive outcomes in aging populations.

Future research should assess whether long-term lifestyle interventions can reduce dementia and SCD risk, as demonstrated in FINGER (Finnish Intervention Study to Prevent Cognitive Decline and Disability) [53]. However, showing their effectiveness in preventing clinically diagnosed dementia remains challenging. AI-driven tools could help by modeling long-term intervention effects, integrating biomarkers to improve risk prediction, and supporting personalized prevention strategies. Additionally, the HHS score should be validated in more diverse populations, and other potentially modifiable risk factors of SCD should be further explored.