Results

Our risk model

We’d be happy to walk you through the results and discuss what an evaluation could look like at your hospital.

What we evaluated

Our research model used prior patient records and information from the first 24 hours of a hospital stay. We tested how well it predicted whether malnutrition would be coded at discharge.

Model performance

The research model achieved 0.93 AUROC on 65,688 held-out admissions. AUROC measures how well the model distinguished between admissions with and without a malnutrition code at discharge. On this scale, 0.5 represents chance-level performance and 1.0 represents perfect separation.

Study scale

We split adult ward admissions from MIMIC-IV into training, validation, and test sets.

Total admissions: 326,222
Training: 195,678
Validation: 64,856
Testing: 65,688

What the research model weighed

On the 64,856 validation admissions, each bar is that group’s share of mean absolute SHAP. A larger share means the group moved the score more, on average. These shares describe the research model on the validation split. The 0.93 AUROC is measured separately, on the held-out testing admissions.

Day-1 labs23%
Care unit & service16%
Outpatient weight & BMI history14%
Day-1 medication orders11%
Admission & ED context9%
Pre-admission labs9%
Demographics & coverage8%
Prior malnutrition codes7%
Prior admissions2%
Other prior diagnoses2%

Rounded to the nearest percent, so the shares can add to 101. Bar length is that share of the whole.

The hospital product

The reported AUROC comes from our evaluation of the research model on historical MIMIC-IV records. The Filos hospital product is shaped with each hospital’s nutrition team. It is built to work with the record, documentation, and CDI systems they already use, and to update malnutrition risk through the stay.

With early hospital partners, we agree on endpoints and success criteria before a deployment. The MIMIC-IV result is the research evidence that we have chosen to share now. A hospital’s own evaluation is how we judge the product in their workflow.

Methods

Adult ward admissions from MIMIC-IV were divided into 195,678 training, 64,856 validation, and 65,688 testing admissions (326,222 in total).

Inputs are limited to information available in the first 24 hours, including history already in the record at that point: day-1 labs, care unit and service, outpatient weight and BMI history, day-1 medication orders, admission and emergency department context, pre-admission labs, demographics and coverage, prior malnutrition codes, prior admissions, and other prior diagnoses.

The label is malnutrition recorded in discharge codes. The model is a gradient-boosted decision tree model (LightGBM). We report AUROC on the held-out testing admissions.

Group shares use mean absolute SHAP on the validation split. Each group’s share is its mean absolute SHAP divided by the sum across groups. The underlying scores are LightGBM log-odds.

Limits of this evaluation

The admissions come from MIMIC-IV, a public dataset from one health system, Beth Israel Deaconess Medical Center. Performance at another hospital can differ.

The outcome follows discharge coding. Malnutrition that was recognized clinically and left uncoded is absent from the label, and coding practice varies.

AUROC describes how the score ranks admissions with a malnutrition discharge code against those without one. Counts of patients a team would review, and estimates of time saved, will be added once they are calculated.

The research model reads the first 24 hours. The hospital product is designed to keep updating risk through the stay.

Data and citations

We evaluated our research model using de-identified hospital records from MIMIC-IV. PhysioNet provides access to credentialed researchers who complete training and sign a data use agreement.

The Malnutrition Screening Tool (MST) curve on our homepage is based on a separate study by Morris and colleagues, who evaluated the MST against the Subjective Global Assessment in acute medical inpatients.

Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, Pollard TJ, Hao S, Moody B, Gow B, Lehman LH, Celi LA, Mark RG. MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data. 2023;10:1. doi:10.1038/s41597-022-01899-x
Pollard T, Moody BE, Lehman LH, Gow B, Fernandes C, Xie C, Johnson A, Mark RG, Heldt T. PhysioNet as a global platform for biomedical research. Nature Health. 2026;1(8):792–795. doi:10.1038/s44360-026-00096-z
Morris NF, Stewart S, Riley MD, Maguire GP. A comparison of two malnutrition screening tools in acute medical inpatients and validation of a screening tool among adult Indigenous Australian patients. Asia Pacific Journal of Clinical Nutrition. 2018;27(6):1198–1206. doi:10.6133/apjcn.201811_27(6).0005

Request the research code

We’re happy to share the code behind this evaluation with members of the research community as we work together to improve care for patients. Contact the founders to request access.

Request code access