Precision Subclassification of Mental Health in Diabetes: Digital Twins for Precision Mental Health to Track Subgroups

People with diabetes (PWD) have a higher risk of mental disorders: a systematic review showed the odds of 15 different mental disorders is significantly increased (1). A review indicated that mental health disorders can drive the incidence of diabetes and lead to a negative prognosis (2). Common mental health problems in diabetes include diabetes distress, depression, anxiety, and disordered eating (1,3).

The ADA/EASD's precision medicine consensus statement highlights mental health as "highly relevant" (4). Moreover, mental health data should be "combined with other data … to improve the precision of clinical decision making." A review by the investigators supports that precision mental health in diabetes can be achieved by combining glucose data with behavioral and mental data (5).

A precision medicine approach to mental health requires an in-depth exploration of mental health subgroups. This involves identifying distinct patterns and subgroups who show different courses of mental outcomes as well as understanding the contribution of glucose in relation to physio- and psychological factors. The interplay between mental health, CGM metrics, and behavioral factors is crucial for elucidating mental health trajectories and advancing personalised treatment strategies (5-7).

With the previous DIA-LINK and PRO-MENTAL studies, the investigators set the basis for a precision mental health approach. By combining CGM with ecological momentary assessment (EMA), a methodology allowing repeated daily assessments of mental variables, an innovative approach for precision mental health was established (7). In a recent study, this approach was used to identify individual drivers of diabetes distress using n-of-1 analyses (8). The study supports specific subgroups in which distress is differentially influenced either by the mental perceptions of glucose or the actual glucose values. It was also demonstrated that glycemic control and psychosocial well-being at follow-up were differentially influenced depending on the individual drivers of distress.

Digital twins are a valuable tool for identifying relevant subgroups of PWD who share similar courses of glucose and mental health. A digital twin represents a virtual model of an individual, integrating data from various sources, including CGM, electronic health records, EMA, behaviors, and sensors. A digital twin is ideal to study highly dynamic networks between metabolic, behavioral, and psychological factors (9-11). However, a recent systematic review highlights that although psychosocial factors play a role in glucose control, none of the reviewed digital twins incorporated such factors. The authors conclude that "addressing this deficiency opens new perspectives and opportunities for improving the holistic management of T1D through more comprehensive and inclusive modeling approaches" (9).

With a digital twin, various treatment scenarios can be simulated. This can help diabetologists to develop precise treatment plans addressing both metabolic and psychological needs, ensuring more personalised interventions. Thus, the present project aims to develop a digital twin for precision mental health in diabetes.

Over the past years, a theoretical framework (5,6) and groundwork for a precision mental health approach in diabetes (7,8,12) have been developed. In the PRO-MENTAL study, a longitudinal panel of over 1,300 PWD who regularly complete a 14-day EMA period and questionnaires every 6 months, in addition to using CGM das been established (12).

The objective of the new project is to capitalize on the achievements of the PRO-MENTAL study and to continue the data collection, while also expanding the database by including the following assessments and data:

• repeated clinical interviews to diagnose mental disorders (depression, anxiety, eating disorders) at follow-up,
• longitudinal data on the prevalence, incidence and remission of key mental disorders,
• CGM parameters and HbA1c,
• person-reported outcomes (PRO) using validated questionnaires (e.g., well-being, diabetes distress, depression, self-management, personality factors, coping, resilience, social support),
• medication doses (e.g., insulin, GLP-1),
• acute and late complications,
• dietary behavior,
• physical activity. By combining these data, the investigators will establish subgroups of people with distinct gluco-psycho-behavioral profiles and map the different trajectories of these subgroups. This will be used to develop a digital twin.

The digital twin can be used to increase scientific understanding of the complex interplay between glucose and psychosocial factors. Previous research on this interplay has mainly focused on specific parts (13-15), but are lacking a holistic approach. A digital twin enables a more comprehensive analyses of subtypes in which psychosocial factors (e.g., stress, mood, diabetes distress, sleep, mental disorders) and different parameters of metabolic control (e.g., hypoglycemia, time in range) influence each other over time. The use of machine learning/artificial intelligence has the potential to reveal complex networks by considering a multitude of influencing and confounding factors simultaneously.

With the digital twin, different courses of mental health (e.g., depression) and glucose control (e.g., time in range) can be simulated. Simulation will be based on different EMA and glucose profiles, PROs, inflammation status, medication, dietary behavior and physical activity. The investigators also aim at simulating how mental health/glucose control will develop depending on changes in the aforementioned variables. By categorizing a person according to their EMA and CGM profiles (8), the investigators aim to predict whether a person is more or less likely to develop a mental disorder or dysglycemia. The digital twin will also allow the simulation of how different interventional measures may affect the overall course of mental health and glucose. This enables the identification and selection of optimal treatment strategies to positively affect both mental health and glycemia. The ultimate goal is to inform clinical decision-making, thereby translating research into precise clinical diabetes care.

Current work on precision mental health in diabetes is entirely preliminary scientific work with no translation to clinical practice. The development of a digital twin would be the first step of a precision mental health approach in diabetes. It could be used by diabetologists in clinical practice as a clinical decision support system to:

• better understand the effect of glucose on psychosocial factors and vice versa within a person,
• identify and prevent critical events in metabolism and mental health,
• support treatment decisions based on simulations,
• inform referrals to psychotherapists when simulations reveal a high probability for the incidence of mental disorders.

Current guidelines, such as the ADA Standards of Care and the German Psychosocial Guideline, recommend monitoring psychosocial factors (16,17). However, guidelines differ significantly in frequency, dimensions assessed, screening tools, and management of positive results. These discrepancies stem from limited understanding of the natural progression of mental health issues in diabetes, the impact of glucose control and complications on mental health, and the implications of these issues for diabetes care. By enhancing our understanding of mental health trajectories in diabetes and the impact of glucose regulation, this study will inform clinical care practices and guideline development.

The TwinPeaks project will be a prospective, longitudinal, observational, non-interventional study including 1,809 participants overall, with ultimately approx. 6,000 person-years being surveyed.

Work packages:

The project will build directly on the PRO-MENTAL study and expand the previous work with five additional work packages (WP).

WP 1 - Data collection for the development of the digital twin: 1,300 participants with type 1 and type 2 diabetes have already been recruited. The majority have completed a baseline diagnostic interview and are attending online surveys as well as a 14-day EMA periods every 6 months. With the TwinPeaks project, the data collection will be continued but more participants and new variables will be included.

1. diagnostic interviews: The diagnosis of a mental disorder is a primary outcome of the digital twin. To train the digital twin how to detect factors that drive the incidence or remission of a mental disorder, an additional clinical interview after baseline is required. Further, for the validation of the digital twin (WP 4) another interview will be performed in a subsample. The diagnostic interviews are conducted by trained psychologists based on the validated Diagnostic Interview for Mental Disorders Mini-DIPS Open Access (18).
2. Metabolic and clinical data: Data from participants' CGM will be collected every 6 months. Information on HbA1c, acute complications (e.g., hypoglycemic episodes, ketoacidosis), and long-term complications (e.g., retinopathy, neuropathy, nephropathy) will be collected.
3. Mental health data: 14-day EMA periods as well as online surveys are conducted every 6 months. The EMA phase will focus on stress, mood, and diabetes distress on a daily basis. The online surveys contains validated questionnaires on different aspects of mental health such as symptoms of depression, anxiety, eating disorder, diabetes-specific fears, well-being, resilience, and coping factors.

WP 2 - Data management and data preparation: The first step is to clean and prepare the plethora of data from different sources and in different formats for further statistical analysis. A key aspect will be the formatting of the data to account for the different time frames: minutes for CGM data, hours and days for EMA data, months for questionnaire data, and years (12 month and lifetime) for clinical diagnostic interviews. The data from several different sources will be made available through a parameterized interface for data collection and storage. Data-centric solutions will be used as the backend, such as employing a NoSQL database. The NoSQL database allows for the storage of unstructured data from various sources and can be horizontally scaled to handle a growing volume of data. A possible implementation is based on Apache Cassandra, known for its high availability and scalability. Suitable metadata and data standards are defined to ensure high comparability. The infrastructure for data collection and storage follows the recommendations of the German Ethics Council for the use of Big Data in medicine.

WP 3 - Development of the digital twin: Digital twins for subgroups are developed for identifying critical events and starting points for intervention from multimodal data using machine learning and data science techniques. Ultimately, optimized approaches will be provided for the automatic identification of patterns, similarities in the data, and outliers. To develop the digital twin, a hybrid modelling approach is used as the mathematical model. This hybrid model leverages two approaches: 1.) Mechanistic models that are based on mechanistic physiological models of glucose regulation. These models are using ordinary differential equations (ODEs) to simulate the regulation of glucose in PWD. 2.) Long-Short Term Memory (LSTM) models that are exclusively data-driven and are more useful when mechanistic physiological models are not available: e.g., there is no physiological model for the impact of mood, diabetes distress or depression on glucose. Both approaches allow the modelling of dynamic systems, but LSTMs are powerful in data-driven, complex and non-linear scenarios, while ODEs perform well in accurately modelling physiological states. The LSTM can be used to model psychosocial factors that are difficult to formalize, whereas ODEs are used to model glucose control, which is well understood and described by physiological laws. Therefore, the Digital Twin uses ODEs to model glucose trajectories and LSTM to capture psychosocial influences. By combining these two methodologies, a digital twin can be constructed that accurately captures complex relationships from a large data set, while adhering to physiological constraints. For the twinning procedure, an appropriate technique will be chosen based on the data and data fit (e.g., Markov Chain Monte Carlo (MCMC)). Taken together, the combination of ODE and LSTM can be used to learn complex, non-linear relationships in the data and make predictions. MCMC can then be used to quantify the uncertainties of these predictions and calibrate the parameters of the ODE/LSTM model. By combining ODE/LSTM networks for modelling temporal dynamics and MCMC methods for uncertainty quantification and parameter calibration, the accuracy and robustness of digital twins can be improved.

With this approach, anomaly detection will be possible, for example, to identify specific behavioral patterns or characteristics associated with an increased risk for mental disorders or critical metabolic events.

WP 4 - Validation process: Validation of the digital twin is a crucial element that will determine the future clinical usefulness. Thus, new data that serves as ground truth for the predictions and simulations made by the digital twin will be collected.

First, it is important to note the kind of predictions we are aiming for with the digital twin. For glucose, we aim at predicting individual glucose control on an aggregate level such as time in range, time below range, coefficient of variation (CV) or HbA1c within a 6-month timeframe. For mental health, we also aim at predicting individual mental health within a 6-month timeframe on a more general level such the course of (sub)clinical mental disorders, questionnaire scores of e.g. well-being, time with distress, or overall stress level.

For validation of the prediction of glycemic control, CGM and HbA1c data will be continuously assessed every 6 months for the whole sample. Similarly, for the validation of the prediction of mental health, EMA will be continuously assessed every 6 months. For feasibility reasons, clinical diagnostic interviews will only be collected in a subset of the sample. Sample size: Usually, for machine learning models, the ratio between training and test data sets is 80:20 (19). Therefore, for validation purposes, a random subsample of 20% of the full sample (-> 0.2 x 1,809 = 362) will be selected for follow-up interviews.

Based on the validation data, the accuracy (e.g., sensitivity, specificity, positive/negative predictive value) and precision (e.g., difference between simulated time in range and actual time in range) of the predictions from the digital twin are assessed.