Clinical natural language processing (NLP) increasingly relies on electronic health record (EHR) data to detect suicidal behaviors, treating clinical documentation as more reliable ground truth than social media posts. However, a new paper from arXiv — authored by Garg, Priyanshi; Rao, Ishita; Ding, Jieqiong; and Paullada, Amandalynne — argues that this framing obscures how EHR-based suicidality datasets encode a particular operationalization of suicidality, shaped by who authors the data, how episodes are bounded, and how ambiguity is resolved.
The paper grounds this argument in a case study of the ScAN dataset, built over MIMIC-III clinical notes. The authors show that the dataset’s construction — including governance constraints, ICD-based cohort selection, single-annotator labeling, and hospital-stay-level aggregation — produces labels that reflect clinician-documented judgments, treat suicidality as a bounded episode, and assume that intent can be reliably inferred from documentation.
The ScAN Dataset and Its Construction
According to the paper, the ScAN dataset’s design choices introduce specific biases:
- Governance constraints: Institutional policies limit what data is available and how it can be used.
- ICD-based cohort selection: Patients are included based on ICD codes for suicidal behavior, which may miss cases not coded.
- Single-annotator labeling: Only one annotator labels each clinical note, introducing subjectivity without inter-rater reliability checks.
- Hospital-stay-level aggregation: Suicidality is treated as a property of the entire hospital stay rather than individual notes, potentially smoothing over temporal nuances.
These factors, the paper argues, encode a narrow view of suicidality that may not reflect the full clinical picture.
Linguistic Heterogeneity Under Identical Labels
The authors conducted a linguistic analysis of the ScAN dataset, revealing that identical labels subsume heterogeneous clinical framings. They found differences in:
- Temporality: Some notes describe past suicidal ideation, others current risk.
- Negation: Phrases like "no suicidal ideation" or "denies suicide" are treated as negative, but clinical context may vary.
- Uncertainty: Use of hedging language (e.g., "possible suicide attempt") creates ambiguity that is collapsed under a binary label.
This suggests that models trained on such datasets may learn spurious correlations rather than true clinical indicators.
Implications for Healthcare AI Deployments
For CTOs and digital transformation leaders in healthcare, the paper's findings underscore the importance of scrutinizing dataset construction before deploying AI models. The assumption that EHR data provides objective ground truth is challenged by the demonstrated biases. Enterprises adopting clinical NLP should demand transparency in:
- Annotation protocols and inter-rater reliability.
- Cohort selection criteria.
- How ambiguity and temporality are handled.
| Bias Source | Effect on Labels |
|---|---|
| Governance constraints | Limits data scope and generalizability |
| ICD-based selection | Misses cases without coded diagnosis |
| Single-annotator labeling | Introduces subjective bias |
| Hospital-stay aggregation | Masks temporal variation |
The paper concludes that clinical NLP should examine the assumptions embedded in suicidality datasets before interpreting their labels as ground truth. This call for rigorous data governance is relevant for any enterprise deploying AI on sensitive clinical data, as dataset artifacts can lead to misclassification and real-world harm.
As research in healthcare AI matures, the lesson from this study extends beyond suicidality detection: the quality of AI systems depends not only on model architecture but equally on the often-overlooked choices made during dataset construction. Enterprise technology buyers should prioritize vendors that can document and justify their data provenance and labeling practices.