Dynamic Risk Prediction for Evolving Clinical States

67 concepts54 papers1 proposal

Traditional clinical risk assessments often function like a snapshot: they capture a patient's status at a single moment, such as hospital admission or the time of diagnosis. However, in acute and chronic care, a patient's condition is a moving target. Dynamic risk prediction shifts the paradigm from static scores to continuous monitoring, updating survival probabilities and complication risks as new data—lab results, vital signs, and even the frequency of medical tests—streams in.

From Snapshots to Trajectories

The core intuition behind dynamic risk prediction is that the trend of a biomarker is often more informative than its absolute value at any single point. For example, Virdee et al. demonstrated that longitudinal trends in full blood counts can identify colorectal cancer risk up to two years in advance, capturing subtle physiological shifts that a one-off test would miss.

This longitudinal approach is equally vital in medical imaging. While AI can identify malignancy in a single CT scan, Paez et al. showed that tracking the trajectory of AI-generated malignancy scores across multiple scans significantly improves the classification of pulmonary nodules. By incorporating the "history" of a nodule, clinicians can better distinguish aggressive cancers from stable, benign growths.

Mechanisms of Real-Time Updating

To turn these trajectories into actionable insights, researchers employ several computational strategies to update risk profiles:

  • Landmarking: This technique involves reassessing a patient’s prognosis at specific time points (landmarks) by incorporating all information available up to that moment. Yan et al. note that landmarking provides a transparent way to address "calibration drift," where static models become less accurate as time passes from the initial assessment.
  • Recurrent Architectures: For high-frequency environments like the ICU, models must process continuous streams of data. Kim et al. utilized Gated Recurrent Units (GRU) to provide continuously updating risk scores for in-hospital cardiac arrest, allowing the model to react to the cumulative and fluctuating effects of vital signs in real-time.
  • Joint Modeling: Some frameworks simultaneously model the underlying health process and the occurrence of events. Futoma et al. developed a scalable joint model that refines individual-specific latent variables as new observations arrive, enabling real-time adjustments to predictions for chronic kidney disease progression.

The Signal in Clinical Behavior

One of the most innovative facets of dynamic prediction is the realization that clinical actions themselves are data points. Alaa et al. introduced the concept of "informative sampling," arguing that the timing of medical tests is a vital signal. Clinicians order tests more frequently when they suspect a patient is deteriorating. By modeling these test timings using Hawkes processes, the authors achieved a significant improvement in deterioration alerts, effectively "learning" from the collective judgment of the medical staff.

Adapting to Treatment Phases

Risk factors are not only time-varying but also phase-dependent. A predictor that is critical immediately after surgery may become irrelevant months later. Hogan et al. found that while early readmissions after kidney transplantation are driven by a mix of clinical factors, the "Length of Stay" from the initial surgery becomes the dominant predictor for late-stage readmissions.

Similarly, in oncology, Dong et al. advocate for dynamic risk stratification that re-evaluates patients based on their biochemical and structural response to initial therapy. This allows for a "risk-adaptive" treatment strategy, where the intensity of follow-up is dialed up or down based on the patient's evolving response rather than their baseline stage.

Clinical Utility and Resource Allocation

The ultimate goal of these models is to support decision-making during surges or in resource-limited settings. During the COVID-19 pandemic, Yao et al. and the authors of the CovEWS system showed that dynamic models could predict mortality risk several days in advance. By reacting to fluctuations in lab results as they occurred, these systems provided a more accurate triage tool than static admission scores, helping doctors prioritize intensive care beds for those whose health trajectories were most precarious.

Key Foundational Work

Predicting the Risk of Complications in Coronary Artery Bypass Operations using Neural Networks (1994) — contributed 1 concept elements to this direction.
Learning Patient-Specific Cancer Survival Distributions as a Sequence of Dependent Regressors (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_risk_assessment_dynamic_risk_pre_a6ab2258,
  title = {dynamic risk assessment (dynamic risk prediction)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-risk-assessment-dynamic-risk-prediction/rdr_a6ab225853904a5dbcdc83745c12ebb8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_risk_assessment_dynamic_risk_pre_a6ab2258_v1,
  title = {dynamic risk assessment (dynamic risk prediction)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-risk-assessment-dynamic-risk-prediction/rdr_a6ab225853904a5dbcdc83745c12ebb8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake