Denial prediction from admission notes
A language model trained on a health system's own clinical notes that reads each inpatient admission's history and physical note and flags encounters whose claims are likely to be denied, so utilisation review and clinical documentation staff can check them during the stay.
Discovery, not a referral queueTechnique: Language Model Text Classification
This approach fine-tunes a language model on the health system’s own clinical notes so it can read the history and physical note written at admission and estimate whether the eventual claim will be denied on first submission.
Its output is a list of flagged inpatient encounters for utilisation review nurses and clinical documentation specialists, who check whether the authorisation is in place and whether the documentation supports the stay while the patient is still in hospital.
The strongest evidence for it comes from NYU Langone Health, where a model pretrained on the system’s own notes predicted initial denials at admission with an AUC of 87% on past encounters.
That result is promising but narrow: it comes from one health system with its own model and team, the denial model was never tested in live use, and the structured model it was compared against was missing fields such as the prior authorisation number.
Using its flags to steer utilisation review during the stay is this playbook’s judgement about where such a model would fit, not a use any source has evaluated.
This is a discovery tool for a large health system that already runs pre-submission checks and root-cause work queues and has the data, the security controls, and the team to train and maintain a clinical language model.
It is the wrong choice for most hospitals and every small practice, where the pre-submission checks and a structured risk score will prevent more denials for far less cost.
Strengths for this problem
- Reads the clinical narrative, which carries information about severity and medical necessity that structured claim fields do not.
- Scores risk at admission, before a claim exists, which in this playbook's judgement is early enough for documentation and authorisation to be checked during the stay.
- In a retrospective evaluation at one health system it predicted initial denial at admission with an AUC of 87%, above a structured model built on claim form fields.
Limitations for this problem
- The published evidence comes from a single large academic health system that pretrained its own model on its own notes, was evaluated only on past encounters, and has not been shown to transfer elsewhere.
- No source shows that acting on a denial flag during the stay prevents the denial.
- The structured comparison in that study lacked several claim fields, including the prior authorisation number, so the size of the advantage over a complete structured model is unknown.
- Needs a large volume of notes, labelled outcomes, and an in-house machine learning team, which most providers do not have.
- A flag says the encounter is at risk, not why, so a person still has to read the note to decide what is missing.
Data needed
- History and physical notes at admission, and discharge notes where the use is later in the stay.
- The insurer's first response to each inpatient claim, linked to the encounter, as the training label.
- Insurer and plan per encounter, so performance can be checked per insurer.
- A secure environment inside the health system where notes can be used for training and inference.
How it is measured
- Share of flagged encounters where utilisation review or clinical documentation staff found a gap in documentation or authorisation.
- Inpatient initial denial rate for flagged encounters that were acted on, against comparable encounters that were not flagged.
- Model performance on a recent time period and per insurer, not only on a random split of past notes.
Pitfalls
- Treating a note-based risk flag as a medical necessity judgement, when the decision belongs to a clinician.
- Prompting clinicians to add documentation to avoid a denial rather than to describe the care, which creates a compliance risk.
- Training on notes and labels from one period and scoring a later one after insurer rules have changed.
- Sending notes to an outside model service without a business associate agreement or without limiting the data sent.
Human oversight
Utilisation review nurses and clinical documentation specialists review each flagged encounter and decide what, if anything, is needed. Physicians decide what is documented about the care, and nothing is added to a note to suit an insurer. The model only flags encounters, and never contacts an insurer, changes a code, or submits a claim.