DPO is a conceptual framework that delineates the core AI functions necessary for DTx to achieve dynamic personalization and optimal therapy. It integrates and analyzes patient data to refine treatment strategies in real time. DPO tailors treatment plans by incorporating patient data and responses to offer individualized therapeutic approaches. A predictive AI model anticipates the patient’s future status and optimizes treatment accordingly. DPO dynamically optimizes personalized therapy by integrating clinical data, feedback, and treatment content.

Building on just-in-time adaptive interventions [10], DPO formalizes relevant data types and introduces adaptive treatment content optimization through AI-based prediction of patient feedback or status. While traditional just-in-time adaptive interventions primarily focus on determining the timing of interventions, DPO differentiates itself by continuously optimizing treatment content through AI-based prediction of patient feedback, thereby extending the concept from temporal adaptation to content-level personalization.

Neurological and other complex disorders require personalized, responsive treatment. DPO is essential to achieving this in DTx, as it continuously tailors therapy based on patient-specific data in real time.

DPO uses four types of data to tailor therapy. User data (U) include information about the patient, such as demographics, medical history, and lifestyle. Status measurement (M) data refer to objective health metrics, including symptom severity and test results. Treatment content data (C) encompass information related to the therapy itself, such as descriptions, types, and formats. Finally, feedback data (F) consist of feedback received after therapy, such as treatment game scores or the patient’s reported feelings. AI in DTx uses these 4 types of data to create personalized treatment plans. Figure 1 illustrates how these 4 data types are sequentially processed throughout the therapeutic process.

As Figure 2 illustrates, instead of directly predicting the optimal content C, this approach focuses on predicting F for a given C. In a DTx environment, real-world data collection yields sequential data (U, M, C, and F) throughout the treatment. An AI model can be trained on these historical data sequences to predict subsequent F. Although F does not directly represent treatment efficacy, it informs future treatment decisions. This approach assumes that feedback (F) is a valid proxy for treatment efficacy. Prior research shows a strong correlation between user engagement in digital interventions and improved treatment outcomes [11], though engagement alone may not fully capture clinical efficacy. Nonetheless, leveraging this correlation, the model can predict F for candidate treatment contents and select the optimal one (C_o) with the highest predicted F.

Figure 3 illustrates another approach: predicting the patient’s M_f. In this approach, an AI model is trained with prior data sequences as input and M_f as the output. If the model learns this relationship accurately, it can predict M_f for a given input sequence and a candidate C. By predicting M for all candidate C options, the model can select the C_o that most effectively improves M.

Together, these approaches eliminate the need to explicitly label C as the correct answer. Instead, AI predicts F or M_f to dynamically determine the optimal C, ensuring that patients engage with the most effective therapeutic content. However, implementing these AI-driven strategies in practice requires managing complex and heterogeneous DTx data, which poses significant challenges.

In a cognitive intervention context for mild cognitive impairment (MCI) [12], DPO operates as follows. The system first collects U such as demographic and behavioral information including age, education level, daily activity, and sleep patterns. It also gathers initial measurement data (M_i), such as the Mini-Mental State Examination score [13] representing the patient’s baseline cognitive state. Based on these inputs and the history of previously delivered C—for example, cognitive games that train memory, attention, or problem-solving skills—and corresponding F, such as game scores reflecting engagement and task performance, the AI model predicts either the expected F or the postsession M for each candidate content option. The predicted value then guides the system in selecting the next C_o for the patient’s subsequent training session. This process of providing C and receiving F repeats over time, allowing the system to refine personalized cognitive training. Through this iterative mechanism, DPO continuously optimizes therapeutic content to enhance cognitive function and engagement in patients with MCI.

DTx data appear in diverse forms—text (eg, clinical notes), numeric measurements, images, and video—varying widely in length and quality. This heterogeneity hinders uniform processing of data by AI. Addressing this issue requires careful data preprocessing, stringent quality control, and AI models capable of integrating disparate data types.

LLMs have significant potential to overcome data processing challenges in DTx [14]. The following strategies describe how LLMs can support DPO.

LLMs can process diverse data formats (including text, numerical data, and images) thereby enabling comprehensive analysis to improve treatment outcomes [15]. They analyze unstructured text, such as medical records and patient feedback, to extract key details and interpret numerical data like vital signs and laboratory results. Additionally, LLMs use embedding techniques to represent each data modality as a numeric vector, enabling integrated analysis across different data types. This approach allows processing of disparate data types within the same mathematical space.

LLMs can handle data inputs of widely varying lengths [16,17]—from brief notes and detailed medical records to extensive patient histories and multisource health logs. They can also process time-series data to identify patterns and trends in symptom logs and vital signs. This capability enables real-time monitoring and more adaptive care.