AgreementHeld-out cases
Agreement with experts90.5%
0%100%
How often the software’s decision matched an expert clinician’s, on cases it had not seen.
Kartini Clinic and cliexa tested decision-support software on records from 606 children and teens with anorexia nervosa. On cases it had not seen before, the software’s decisions matched expert clinicians’ decisions 90.5% of the time. The paper is in press at PLOS Digital Health.
How often the software’s decision matched an expert clinician’s, on cases it had not seen.
On a scale where 0.5 is chance and 1.0 is perfect (AUC-ROC).
The diagnosis was classified correctly in 79 of 86 cases.
Children and teens ages 7 to 17 with anorexia nervosa.
Source: O’Toole J, Bermudez O, Kazgan M, et al. Clinical decision support in pediatric anorexia nervosa. PLOS Digital Health, in press.
Kartini Clinic and cliexa tested decision-support software on records from 606 children and teens with anorexia nervosa. On cases it had not seen before, the software’s decisions matched expert clinicians’ decisions 90.5% of the time. The records also showed that age and early weight gain are strong signs of how treatment will go.
Kartini Clinic is a specialty pediatric eating disorder treatment center in Portland, Oregon, founded in 1998 by Dr. Julie O’Toole.
The clinic treats children and adolescents ages 6 to 18, with a medical-first, family-centered, evidence-driven approach to care.
Researchers at Kartini Clinic used cliexa to study treatment records from 606 children and teens, ages 7 to 17, with anorexia nervosa. The records came from four clinical datasets collected over 20 years, from 2000 to 2020.
In plain terms: the team looked for patterns in past patient records, used those patterns to build decision-support software, and then tested the software against the decisions of expert clinicians.
First, standard statistics looked for patterns in the records. They found three:
Then cliexa’s clinical rules engine turned those patterns into inputs for a machine learning model (a type called gradient boosting, or XGBoost). It also pulled out a set of simple, readable rules (using a method called RIPPER) that a clinician can read and check.
The software was tested on cases it had not seen during training, and its decisions were compared with those of expert clinicians.
How closely it matched expert doctors
On the cases held back for testing, the software’s decision matched the expert clinician’s decision 90.5% of the time.
This score (called AUC-ROC) shows how well a model tells outcomes apart. 0.5 is no better than chance and 1.0 is perfect. The likely range here is 0.81 to 0.92.
The software classified the diagnosis correctly in 79 of 86 cases.
A closer look at its accuracy
The records also revealed clear patterns in who responds to treatment, and how quickly.
Age 14 was a dividing line
of this group had the restricting type of anorexia nervosa (AN-R).
of this group had the restricting type (AN-R). The gap is statistically significant, meaning it is very unlikely to be due to chance (χ² = 60.9, p < 0.001).
Younger patients did better
reached 100% of their expected weight within 120 days.
reached 100% of their expected weight within 120 days.
cliexa combines patient context and clinical protocols to guide action within existing workflows. Clinicians retain judgment.
O’Toole J, Bermudez O, Kazgan M, et al. Clinical decision support in pediatric anorexia nervosa. PLOS Digital Health, in press.
Bring a care pathway and we will show you how cliexa combines patient context and clinical protocols to guide action within your existing workflow.