Imagine watching a single frame of a movie. If a car is captured at position $X$, it is impossible to tell whether it is accelerating, braking, or completely stationary. Traditional clinical diagnostics often treat patient health exactly like this: as a series of isolated, static snapshots. As Bjoern Eskofier bars et al. note in Predictive Models for Health Deterioration: Understanding Disease Pathways for Personalized Medicine, relying on single-point-in-time snapshots fails to capture the continuous evolution of chronic or recurring illnesses.
When models only look at baseline measurements, they miss the velocity of change. Two patients might present with identical biomarker values today, but one may be stable while the other is rapidly deteriorating. To build truly personalized medicine, clinical AI must become "temporally aware," shifting the paradigm from static classification to longitudinal trajectory modeling.
Quantifying the Derivative: Velocity Over Baselines
The core reason longitudinal trajectory modeling works is that the mathematical derivative—the rate of change or slope of functional loss over time—contains predictive signals that static thresholds obscure.
In neurodegenerative diseases, for instance, Daoqiang Zhang et al. demonstrated in Predicting Future Clinical Changes of MCI Patients Using Longitudinal and Multimodal Biomarkers that the rate of change in physiological markers is a far more potent predictor of cognitive decline and eventual conversion to Alzheimer's Disease than static baseline measurements. By prioritizing temporal transitions and rates of change over baseline values, models can account for individual variability in progression speed, as shown by Michael C. Burkhart in Unsupervised multimodal modeling of cognitive and brain health trajectories for early dementia prediction.
This principle applies across diverse physiological systems:
- Nephrology: Rather than relying on a single snapshot of kidney function, Lijie Xie et al. calculated the annual rate of change across multiple clinical visits to distinguish stable patients from fast-declining ones in Anemia is a risk factor for rapid eGFR decline in type 2 diabetes. Similarly, Yuqing Chen et al. tracked the velocity of clinical metrics to isolate specific metabolic signatures linked to rapid kidney function decline (Metabolomics profiling in multi-ancestral individuals with type 2 diabetes in Singapore identified metabolites associated with renal function decline).
- Structural Failure: In predicting spinal fractures, calculating the slope of change across multiple intervals reveals impending structural failures before they manifest physically. Leander Glanz showed that a rapid increase in bone marrow fat—measured via the slope of change—serves as a critical warning sign for vertebral compression fractures, even when traditional static bone density measurements appear stable (Associations of incidental vertebral fractures and longitudinal changes of MR–based proton density fat fraction and T2* measurements of vertebral bone marrow). This is reinforced by Yannick Stohldreier et al. in Longitudinal assessment of changes in muscle composition using proton density fat fraction and T2* in patients with and without incidental vertebral compression fractures, where rate-of-change analysis over a defined period captured dynamic shifts in tissue composition that baseline measurements missed.
Dissecting Disease Heterogeneity via Trajectory Clustering
Diseases like Parkinson's or Acute Respiratory Distress Syndrome (ARDS) are notoriously heterogeneous. Cross-sectional snapshots often lump completely different biological pathways together because their symptoms look similar at a single point in time. Trajectory clustering solves this by grouping patients based on the shape and direction of their progression over time.
In neurology, Rainer von Coelln et al. applied bipartite network clustering to disease velocity patterns in Identifying and predicting Parkinson’s disease subtypes through trajectory clustering via bipartite networks. By clustering patients based on their longitudinal patterns of change rather than cross-sectional snapshots, they successfully identified stable, predictable patient subgroups.
Static Snapshot (Time T) Longitudinal Trajectory (Times T1 -> T3)
Patient A (Score: 50) Patient A (50 -> 50 -> 50) ==> Stable Subtype
Patient B (Score: 50) Patient B (70 -> 50 -> 30) ==> Rapidly Declining Subtype
In acute care settings, the authors of Identifying novel clinical phenotypes of acute respiratory distress syndrome using trajectories of daily fluid balance: a secondary analysis of randomized controlled trials discovered two distinct phenotypes of ARDS based on daily fluid balance trajectories. One group struggled to clear fluid, while the other did so efficiently. Lumping these patients together under a single static diagnosis leads to sub-optimal "one-size-fits-all" treatments, whereas trajectory analysis reveals who will actually benefit from specific fluid management strategies.
Similarly, long-term tracking of inflammatory markers like high-sensitivity C-reactive protein (hsCRP) by Jianwei Pan et al. revealed that patients with increasing or high-but-decreasing trajectories faced vastly different risks for Metabolic Syndrome compared to those with stable baseline levels (Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study).
Pinpointing Critical Windows and Inflection Points
Trajectories do not just tell us where a patient is going; they tell us when they will get there. By mapping continuous biomarker curves, researchers can identify critical thresholds and acceleration phases where interventions are most likely to succeed.
During acute infections, tracking the daily progression of biological markers allows clinicians to catch systemic deterioration early. Ming Wang et al. used dynamic monitoring over a fixed temporal window to observe diverging trends in liver enzymes and inflammatory indicators, pinpointing the exact time points where physiological decline accelerated (Clinical characteristics and risk factors of liver injury in COVID-19: a retrospective cohort study from Wuhan, China). Similarly, Tuo Hu et al. mapped daily cardiac biomarker trajectories to identify a critical window between days 14 and 20 where cardiovascular complications peak, establishing a clear timeline for proactive monitoring (Cardiac Biomarker Abnormalities Are Closely Related to Prognosis in Patients with COVID-19).
This temporal mapping applies to non-acute settings as well. In Prevalence and course of pregnancy symptoms using self-reported pregnancy app symptom tracker data, Michael Nissen et al. transformed noisy, self-reported app data into precise longitudinal trajectories of pregnancy symptoms. This uncovered distinct weekly peaks (such as a localized increase in headaches around week 15) that broad trimester-based generalizations completely obscure.
Algorithmic Architectures for Continuous Trajectories
How do we mathematically model these trajectories? Traditional clinical models often treat hospital visits as isolated snapshots rather than a continuous trajectory. Modern frameworks address this by treating hospital visits as observations from an underlying continuous state-space.
For example, Jing Ma et al. developed DSLAM (Deep State-space with Label-dependent Attention Model) in Modelling Patient Longitudinal Data for Clinical Decision Support: A Case Study on Emerging AI Healthcare Technologies. This framework integrates state-space models with deep learning to capture the time-varying progression of chronic diseases across multiple hospital visits. To align behavioral shifts with clinical outcomes, Robert Lewis et al. introduced a two-step slope analysis that correlates the long-term rate of change in digital biomarkers (like speech patterns) directly with the velocity of clinical recovery or decline (Towards the Objective Characterisation of Major Depressive Disorder Using Speech Data from a 12-week Observational Study with Daily Measurements).
Furthermore, longitudinal trajectory modeling allows researchers to distinguish between immediate, acute responses and long-term maintenance phases following an intervention, as demonstrated by Qiao Gao et al. in Comparative effectiveness of first-line antihypertensive drug classes on the maintenance of estimated glomerular filtration rate (eGFR) in real world primary care. By looking backward at historical operational data, frameworks like the retrospective electronic frailty index developed by Tabinda Sarwar et al. can reconstruct a patient's physiological decline over time, providing actionable insights that a single static assessment cannot replicate (Development and validation of retrospective electronic frailty index using operational data of aged care homes).
By shifting focus from what a biomarker value is today to how fast and in what direction it is moving, longitudinal trajectory analysis unlocks pre-diagnostic signals (Ana I. Robles et al., Excerpts from the 1st international NTNU symposium on current and future clinical biomarkers of cancer: innovation and implementation, June 16th and 17th 2016, Trondheim, Norway) and long-term pathological shifts (Isao Ishikawa, Renal cell carcinoma in chronic hemodialysis patients-A 1996 questionnaire study.). It transforms static data into a dynamic vector, offering a clearer window into the true velocity of human disease.
Key Foundational Work
An Epistemology for Clinically Significant Trends (1993) — contributed 1 concept elements to this direction.
Renal cell carcinoma in chronic hemodialysis patients-A 1996 questionnaire study. (1998) — contributed 1 concept elements to this direction.