In supervised learning, we treat labels as ground truth. We assume that if a patient is labeled with "Sepsis" at 2:00 PM, the underlying biological reality of sepsis began at 2:00 PM. However, in clinical practice, a label is rarely a timestamp of biological onset; it is an administrative artifact of documentation, laboratory confirmation, or physician observation.
This discrepancy creates a temporal gap—a systematic delay between the physiological start of a condition and its appearance in a dataset. Research across diverse medical domains suggests that failing to account for this lag leads to models that are biologically observant but statistically penalized.
The Latency of Ground Truth
The most pervasive issue in medical event data is the reliance on diagnosis timestamps. As Zhaozhi Qian, Ahmed Alaa, and Mihaela van der Schaar argue in their work on comorbidity networks, chronic diseases often begin months or years before they are formally documented. When we train models on Electronic Health Records (EHR), we are often modeling the "age at diagnosis" rather than the "age at onset."
This isn't just a matter of precision; it’s a fundamental data gap. Xilin Jiang and Alkes L. Price demonstrate that this discrepancy introduces significant noise into longitudinal studies, particularly when trying to identify disease subtypes. If two patients have the same biological condition but one is diagnosed three years later due to healthcare access or symptom variance, a naive model will treat them as having different disease trajectories.
The Early Detection Penalty
A counterintuitive consequence of this lag is that highly sensitive models can be penalized for being "too good." In Alzheimer’s classification, Arjun Punjabi and colleagues found that models fusing MRI and PET scans often identify early pathological signs in patients still labeled as "healthy" by clinicians.
Because the clinical label lags behind the biological marker, the model’s correct identification of early-stage pathology is recorded as a false positive. This "mismatch between biological reality and clinical categorization" means that supervised learning objectives may inadvertently discourage the very early-warning capabilities we seek to build.
Misalignment in Acute Care
The problem intensifies in high-acuity settings like the Emergency Department or Intensive Care Unit, where minutes matter.
- Sepsis: Aaron Boussina notes that ground-truth labels for infection are often tied to the timing of laboratory cultures. However, physiological symptoms manifest at variable intervals relative to when a test is ordered, making it difficult to synchronize real-time clinical data with the eventual label.
- Heart Attacks: Daniel Gedon and Stefan Gustafsson highlight that training on hospital discharge summaries—which reflect the cumulative clinical picture—can cause a model to predict an eventual outcome rather than the current physiological state captured in an ECG.
- Imaging: Even in radiology, a time gap between an initial scan and a confirmatory "gold-standard" test can introduce noise. Isabelle Riederer and Lei Zhu show that if a condition like a stroke or spinal compression progresses during that interval, the ground truth used for training no longer reflects the anatomy captured in the original image.
Modeling the Hidden Timeline
To solve this, researchers are moving away from treating labels as static points in time and toward modeling them as outcomes of a continuous process.
One approach is Age-dependent Topic Modeling (ATM), which Jiang et al. used to uncover hidden disease subtypes by analyzing how comorbidities cluster over a lifetime, rather than at a single diagnosis event. Similarly, the TRADE model uses a Transformer-based foundation model to predict Alzheimer’s risk up to five years in advance, effectively "looking past" the formal diagnosis date to the preceding risk trajectory.
Another strategy involves self-supervision. Arinbjörn Kolbeinsson and Luca Foschini demonstrated that by training models to predict "tomorrow's" health metrics from wearable data, they could detect influenza-like illnesses before a laboratory diagnosis was available. By focusing on the continuity of the signal rather than the discrete label, the model becomes robust to the inherent lag of clinical confirmation.
Ultimately, the goal of this research direction is to transform medical AI from a system that mimics documentation into one that understands the underlying biological clock. By acknowledging that the "ground truth" is often late, we can build models that are early.