Table 1 Examples of EHR-based clinical tasks DL-ClinAI can support

Disease diagnosis

Relies primarily on patient symptoms and test results to determine the presence of disease⁹³.

Improves timeliness and accuracy by integrating clinician actions with patient data to capture context-aware diagnostic signals. For example, patterns like frequent laboratory test review and revisits to the same diagnostic section in the EHR may signal diagnostic uncertainty or ongoing reconsideration. Diagnosing of some complex conditions requires clinicians to sequentially access diverse EHR sections, such that specific navigation sequences themselves may reflect the information gathering process critical to diagnostic reasoning.

Patient phenotyping and subtyping

Categorizes patients based on structured EHR data elements such as diagnosis codes, laboratory values, and medications^94,95.

Enhances phenotyping by incorporating selected and normalized clinician behaviors from the EHR (e.g., patterns in ordering, response strategies, and documentation intensity), which adds critical context to patient data and helps to delineate more nuanced patient subgroups. For example, patients sharing the same diagnosis codes but exhibit more intensive clinician documentation, repeated diagnostic testing, and frequent specialist consultations may represent a clinically distinct or more complex phenotype that would be missed by traditional models. Key strategies, such as stratifying clusters by key sociodemographic factors for detecting confounded patterns and validating emergent phenotypes against independent clinical endpoints, can help mitigate potential biases.

Outcome prediction

Uses objective clinical variables to forecast patient trajectories and critical clinical events^93,96.

Combines patient data with clinician workflow signals to make context-aware risk estimates. Behavior traces do not merely repeat a clinician’s explicit judgment. Rather, they aggregate actions from a care team that is typically more informative regarding a patient’s condition than a single clinician’s action sequences. Behavior traces that occur before overt escalation and before new findings are formally codified in the EHR can serve as early proxies of rising concern within the care team, which enables the model to flag high-risk patients timely. For example, two patients with comparable laboratory results and vital signs may exhibit divergent risk profiles if one triggers immediate clinician responses, such as urgent vital monitoring, note documentation, order entry, and specialist consultation, whereas the other receives only a routine follow-up process. DL-ClinAI can surface high-risk cases sooner and spotlight potential recognition gaps when clinical response lags behind patient need.

Intervention recommendation

Generates suggestions based on clinical indicators⁹⁷.

Enhances recommendations by integrating the context of clinician decision sequences and observed practice patterns, leading to recommendations that are more actionable and aligned with real-world clinical workflows. For example, rather than suggesting an aggressive intervention immediately after a single test result, the new approach may recognize that clinicians typically opt for observation and repeat testing first and thus suggest a more aligned intervention sequence reflective of the context.

CDS alert design and optimization

Employs static thresholds and rules derived solely from patient data, often leading to alert fatigue⁹⁸.

Leverages insights from behavioral data to develop intelligent AI-assisted CDS systems that modulate timing, priority, and routing based on real-time clinician actions and context. The trained model can minimize unnecessary notifications and ensure that high-priority alerts are aligned with real-time clinical urgency. For example, if EHR use metadata indicates that a clinician has already reviewed a patient’s relevant labs, navigated to diagnostic imaging, and documented an assessment, the system can infer active engagement and suppress redundant alerts. Conversely, if inattention or disengagement is detected, the system may escalate the alert to another team member or adjust its urgency. The core mechanisms can be learned through AI-driven modeling of the behavior patterns.

Workflow modeling and optimization

Focuses mainly on observations and patient charts, offering a coarse view of care processes while overlooking the subtleties of clinical processes that drive real-world inefficiencies⁹⁹.

Integrates EHR use metadata reflecting nuanced clinician activities to model workflows more comprehensively, identify bottlenecks, and suggest process improvement. For example, by identifying and analyzing repeated handoffs, delayed task completion, and frequent task switching in the EHR, the new approach can signal communication breakdowns and cognitive overload that traditional models would fail to detect.

Patient triage and prioritization

Uses clinical indicators to rank patient urgency, generally treating all patients with similar clinical profiles equivalently¹⁰⁰.

Incorporates both clinical data and real-time and historical patterns of clinician behaviors to infer latent urgency signals and contextual cues that are not explicitly recorded, achieving more anticipatory and context-sensitive triage and prioritization. For example, a sudden increase in chart access frequency, note revisions, when combined with subtle changes in patient vital signs, can be detected as an indicator of an emerging clinical deterioration.