Every day, a child with a stomachache arrives at an emergency department, and somewhere in that ordinary complaint lies one of medicine's quieter difficulties: appendicitis is hardest to recognize in the youngest patients, who often cannot describe what hurts and whose bodies do not follow the textbook. In a January 2026 paper in PLOS Digital Health, researchers in Germany introduced Dharma, an open-access machine learning tool meant to help clinicians weigh routine clinical, laboratory and imaging data both to diagnose the condition and to judge whether it is becoming dangerous. It is an atte
Dharma: Interpretable Machine Learning Tool Aims to Improve Pediatric Appendicitis Diagnosis and Severity Assessment
Children with suspected appendicitis face risks from negative appendectomy, unnecessary CT radiation, and sedation, with perforation rates above 80% in children under three.
Missing data, in this tool, is treated as a clinical signal rather than a gap.
Mark
Why does a tool like this matter if doctors already have scores like Alvarado?
Mimi
Because those scores are fixed and linear, and the paper says none is accurate enough to stand alone. Dharma tries to capture interactions between features and explain each prediction.
Luke
Worth noting that the excerpt I can see doesn't give Dharma's performance numbers. We know what it was compared against, not how it came out.
Mark
The misdiagnosis figures are striking, 70% to 100% in children three and under.
Mimi
And perforation above 80% in that same youngest group. The stakes of a missed or late call are highest where diagnosis is hardest.
Luke
Those are figures the authors cite from earlier literature, not findings from this study.
Mark
How does it handle missing data?
Mimi
Cleverly, I think. A missing appendix diameter, 36% of cases, is treated as a signal that the appendix wasn't visualized, coded as -1 with a flag.
Luke
But the data are all from German hospitals, Regensburg and Düsseldorf. Claims about resource-limited settings are a design goal, not something demonstrated here.
Mark
So what would settle it?
Mimi
Prospective testing in real clinics.
Luke
Agreed. Past ML tools looked good retrospectively and still weren't adopted.
The Pulse
- Misdiagnosis of appendicitis in children three and under climbs as high as 70% to 100%, leaving clinicians navigating symptoms that rarely present cleanly.
- Negative appendectomies, unnecessary CT scans and sedation burden children who may not have needed them at all.
- Perforation rates above 80% in the youngest patients make the cost of a missed or delayed diagnosis painfully concrete.
- Dharma is being tested on retrospective hospital records from Regensburg and Düsseldorf, with the open question of whether it can prove itself in live, prospective clinical use.
Every day, a child with a stomachache arrives at an emergency department, and somewhere in that ordinary complaint lies one of medicine's quieter difficulties: appendicitis is hardest to recognize in the youngest patients, who often cannot describe what hurts and whose bodies do not follow the textbook. In a January 2026 paper in PLOS Digital Health, researchers in Germany introduced Dharma, an open-access machine learning tool meant to help clinicians weigh routine clinical, laboratory and imaging data both to diagnose the condition and to judge whether it is becoming dangerous. It is an attempt, in the long human effort to read the body more truly, to offer a second pair of reasoning eyes, and its worth will be measured not in its promise on paper but in whether it is trusted at the bedside.
A child arrives at an emergency department with a stomachache. It is one of the most common complaints in pediatric medicine, and appendicitis is among the hardest diagnoses to make, especially in the very young, who often do not show the classic signs. A new open-access tool called Dharma, described in a January 2026 paper in PLOS Digital Health, aims to help clinicians with exactly this uncertainty.
Existing scoring systems, including the widely used Alvarado score, have not proven sensitive or specific enough to stand alone, and more imaging has not closed the gap. The consequences fall on children: unnecessary surgeries where a healthy appendix is removed, radiation from scans, and the risks of sedation. In the other direction, no marker reliably predicts when a simple case will become a complicated one, and perforation is common in children under three.
Dharma draws on routine clinical, laboratory and radiological data, models the interactions between features that fixed scores cannot capture, and explains its predictions to the clinician. It was built from records of 782 children at Children's Hospital St. Hedwig in Regensburg, with external testing on additional cohorts from Regensburg and Florence-Nightingale Hospital in Düsseldorf. Missing data was a central challenge, and the authors handled absent appendix measurements as meaningful signals rather than gaps to be filled.
Machine learning has been tried in appendicitis since the 1990s without wide adoption. Whether Dharma breaks that pattern depends on what comes next: prospective trials, live clinical use, and performance across hospitals with very different resources.
A child arrives at an emergency department with a stomachache. It is among the most ordinary complaints in pediatric medicine, accounting for up to 10% of all emergency visits by children, and in those over age one, appendicitis is the leading cause. It is also one of the harder diagnoses to make. A new open-access tool called Dharma, described in a January 2026 paper in PLOS Digital Health, is an attempt to help clinicians with exactly that problem.
The authors lay out why the problem persists. Children often present without the classic symptoms, and the appendix varies anatomically, including in diameter. The Alvarado score is the oldest and most widely used scoring system, but the authors say none of the existing scores has enough sensitivity, specificity or predictive value to serve as the sole basis for a diagnosis. Heavier use of CT and ultrasound has not brought a matching improvement in detection. Misdiagnosis rates run from 70% to 100% in children aged three or younger and decline as age rises.
The consequences fall on children. Negative appendectomy, in which the appendix is removed and turns out to be healthy, is reported at 5% to 20% in various centers worldwide, and carries the risks of general anesthesia and surgery. Unnecessary CT scans expose children to radiation, and sedation carries its own risks. In the other direction, no traditional inflammatory marker, sign, symptom or imaging method reliably predicts when uncomplicated appendicitis will progress to a complicated case. Perforation is common in the youngest patients, reaching over 80% in children under three.
Dharma is built to support both diagnosis and severity assessment using routinely available clinical, laboratory and radiological data. The authors say it is meant to mirror clinical reasoning rather than replace it, to model non-linear interactions among features, which fixed linear scores cannot do, and to explain individual predictions. It runs as a web-based tool that gives real-time risk estimates and is intended to adapt to different resource settings, including places with limited imaging. The team compared it against conventional tools, including the Alvarado score, the Pediatric Appendicitis Score, the AIR score, the Tzanaki score and ultrasound, and against the machine learning architectures XGBoost and LightGBM.
The data come from public, anonymized datasets. The main cohort covers 782 children admitted with abdominal pain to Children's Hospital St. Hedwig in Regensburg, Germany, between 2016 and 2021; diagnostic labels were available for 780. Of those, 463 had appendicitis and 317 did not, and 119 had complications. Each record held 55 variables. For external testing of the diagnostic model, the team used 239 unique records from a separate 430-case collection from the same hospital covering 2016 to 2018. For the severity model, they added that collection plus 49 complicated cases from a 301-record set gathered between 2015 and 2022 at Florence-Nightingale Hospital in Düsseldorf. After removing duplicates, the combined set held 1,069 patients. Only the 650 with confirmed appendicitis were used for severity modeling: 490 uncomplicated and 160 complicated.
Missing data is a central design issue. Nearly every variable had gaps. The authors built a custom imputation pipeline, Dharma_Imputer, which groups inflammatory markers such as white cell count, neutrophil percentage, C-reactive protein and temperature and fills them in with decision-tree methods. Appendix diameter, missing in 36% of cases, was handled differently. Because an appendix that cannot be seen on ultrasound is common and meaningful in itself, missing values were flagged and set to a placeholder of -1. Imputation was run separately on training, validation and test sets, and the same imputer is built into the web tool. Treatment choice was left out of the analysis, since the authors regard it as the surgeon's decision.
The authors acknowledge that machine learning has been applied to appendicitis since the 1990s, with pediatric models appearing in the early 2010s, and that none has been widely adopted as a real-world decision support system. Whether Dharma breaks that pattern depends on testing beyond retrospective data, in live clinical use, across hospitals with very different resources.
Notable Quotes
Dharma is designed to mirror real-world clinical reasoning rather than replace clinical judgment.— Study authors, PLOS Digital Health
Want the full story?
Read the original at PLOS (Public Library of Science) ↗
Technology