POCUS and Bayes at the Bedside: Is AI Ready to Help Nurses Diagnose Hydronephrosis?

Lauren Breslin, MD; Amiran Baduashvili, MD*

Division of Hospital Medicine, University of Colorado School of Medicine, Aurora, CO, USA

*Corresponding Author: Dr. Amiran Baduashvili (email: amiran.baduashvili@cuanschutz.edu)


Download article PDF – POCUS Journal 2026;11(2):14-16

DOI: https://doi.org/10.24908/pocusj.v11i02.22264


Point of care ultrasound (POCUS) can expand timely access to diagnostic imaging and reduce reliance on radiation-based studies, but its use is often limited by the availability of clinicians trained to acquire and interpret images. Artificial Intelligence (AI)-assisted interpretation of POCUS images for a focused clinical question could help overcome this bottleneck, extending POCUS to settings where trained sonographers or POCUS-proficient physicians are scarce.

In a single-center study from an academic hospital in Japan, Matsumoto and colleagues developed and validated a machine learning image interpretation algorithm to detect hydronephrosis and assess prostate volume from POCUS images, in two phases [1]. First, they trained the algorithm on images from two sources: a professional sonographer using a console ultrasound machine (traditional high-end radiology equipment), and the same point-of-care device later used for prospective data collection. Second, they tested the tool prospectively on images obtained by nurses, who received just 30 minutes of instruction before scanning renal and prostate views in patients admitted to the hospital’s urology floor. One kidney per patient was evaluated. A sonographer, blinded to the AI’s output, independently classified each image for the presence or absence of hydronephrosis (without grading severity) and for a prostate volume exceeding 50 mL, a threshold for severe enlargement. The analysis included 32 kidneys (one per patient) and 28 prostates, after 5 non-diagnostic prostate scans were excluded. Most patients were over 60 years old, had a normal body-mass index (20–25 kg/m²), and about half were admitted for a genitourinary diagnosis.

Among the 32 patients evaluated, 15 (47%) had hydronephrosis confirmed by a sonographer. The AI tool correctly identified disease in 14 of 15 cases (sensitivity 93%) and correctly identified its absence in 14 of 17 patients without hydronephrosis (specificity 82%). The positive predictive value (PPV; the probability of true hydronephrosis if the AI tool reported a positive result) was 82%, and the negative predictive value (NPV; the probability of being hydronephrosis-free if the AI tool was negative) was 93%. The corresponding sensitivity, specificity, PPV, and NPV for severe prostate enlargement were 71%, 71%, 45%, and 88%. The authors concluded that their results support using this AI model to triage the presence of hydronephrosis on nurse-acquired POCUS images, with a more cautious endorsement for prostate volume assessment pending further validation.

The authors envisioned deploying this tool in nursing-home and outpatient settings, where nurses could use AI-assisted POCUS to triage patients with lower urinary tract symptoms who need further evaluation. A precedent already exists in automated post-void residual volume calculators, but those devices cannot identify an obstructive cause like prostate enlargement or detect upstream hydronephrosis. This tool could fill those gaps—letting nurses complete a more thorough evaluation and accelerate care. But, is this tool ready for that kind of real-world deployment?

When Predictive Values Predict Confusion

Imagine a co-author of this editorial claims he can see through people’s bodies and rule out hydronephrosis by sight alone. If true, the savings on POCUS devices, electricity, AI tokens, and ultrasound gel would be substantial. But first, we have to put his claim to the test. We ask him to look at a lecture hall of 100 medical students and guess who has hydronephrosis and who does not. Afterward, we scan every student with ultrasound, have a radiologist read the images, and compare those reads against his “eye test.” Remarkably, our co-author achieves a 99% NPV: 99 out of 100 times he called a student hydronephrosis-free, he was right.

Should we unleash his extraordinary powers at the nearby nursing home? Probably not. It turns out only 1 of the 100 students actually had hydronephrosis, and our co-author simply labeled everyone hydronephrosis-free. His NPV looks impressive only because the prevalence of hydronephrosis was so low. The same trick works in reverse: asked to diagnose heart failure in a cardiac intensive care unit, where most patients already have it, he could rack up an equally impressive PPV just by labeling everyone positive. The lesson is twofold: predictive values are driven heavily by disease prevalence, and an impressive predictive value can make a test look far more useful than it is.

Unlike our co-author, this AI tool does discriminate between kidneys with and without hydronephrosis. The usefulness of a test is better captured by likelihood ratios, which describe how much a given result shifts the odds of disease [2]. The further away the likelihood ratio is from 1, the more it shifts disease probability, or the more useful it is. For hydronephrosis, this AI tool’s positive likelihood ratio (sensitivity ÷ [1 − specificity]) is about 5, and its negative likelihood ratio ([1 − sensitivity] ÷ specificity) is about 0.08. Both likelihood ratios are meaningfully different from 1, unlike our co-author’s, whose negative likelihood ratio is exactly 1, despite his impressive NPV. But before interpreting any predictive value, even for a test with genuine discriminating power like this one, we have to ask whether the study’s prevalence matches the prevalence of the setting where we plan to deploy it.

In the study by Matsumoto and colleagues, nearly half of the patients scanned had hydronephrosis—far higher than we would expect among community-dwelling adults, even those with lower urinary tract symptoms. Data on the prevalence of hydronephrosis are limited. Estimates range from 3.1% in a large autopsy study spanning all age groups, from neonates to adults [3], to 17% among men undergoing urodynamic testing, with a prevalence of 9.5% among men aged 60 years and older in the same study [4]. Assuming a 10% prevalence among community-dwelling adults with lower urinary tract symptoms, and applying this study’s reported sensitivity and specificity, we would expect 100 of 1,000 patients to have hydronephrosis: 93 of those 100 would correctly test positive, and about 740 of the remaining 900 would correctly test negative (with 160 false positives). That leaves 253 patients testing positive: only 93 of whom actually have hydronephrosis, dropping the PPV from an impressive 82% to approximately 36%. We would get the same answer if we converted the pretest probability of 10% to pretest odds of 1/9, multiplied it by the positive likelihood ratio (5), and converted the posttest odds (5/9) to posttest probability (5/14 ~36%).

In sum, predictive values can be misleading. Readers should ask how good the likelihood ratios are at changing the disease odds and whether the disease prevalence in the study is similar to the prevalence in the setting where the tool will be used. Higher-than-expected prevalence can also signal another issue—spectrum bias.

Spectrum Bias and Smoke Detector

Imagine a smoke detector that has only ever been tested in buildings already fully on fire. Unsurprisingly, the alarm will sound every time. Can we conclude this is a very sensitive smoke detector? No, we cannot. A raging fire is easy to detect, but we do not know how well this device will perform with subtler signs of early fire, such as a few wisps of smoke coming from a faulty electric wire. This same idea may apply to the AI tool in this study. It was evaluated on a urology floor, where patients are typically referred because hydronephrosis is advanced—the equivalent of a smoke detector that has only ever seen buildings ablaze. So, the excellent sensitivity reported here may partly reflect how obvious these cases were, rather than how good the tool truly is. In a nursing home or outpatient clinic, where hydronephrosis is often mild and found incidentally, the same tool may struggle far more than these results suggest. In such a population, true sensitivity may be markedly lower than what the authors found. In diagnostic test accuracy literature, the issue of inflated sensitivity due to selection of sicker patients with more obvious disease presentation is known as spectrum bias [5].

Spectrum bias can also affect the test specificity. On a urology floor, the group without hydronephrosis is not made up of patients with healthy kidneys. It is reasonable to assume patients without hydronephrosis have other urologic pathology: renal cysts, masses, and other mimics that can be mistaken for hydronephrosis on ultrasound. The presence of mimics in the group of patients without the disease we are looking for can lead to underestimation of specificity because these mimics may show up as false positives. In a general or broader target population, where the non-disease group mostly has normal kidneys with far fewer confusing mimics, true specificity may actually be higher than this study suggests. Strikingly, both distortions trace back to the same root cause: the referral pattern that pushed hydronephrosis prevalence to 50% simultaneously inflated sensitivity by selecting florid disease and deflated specificity by selecting mimics. We cannot know the degree to which the spectrum bias affects the reported test characteristics, as the authors did not report data on hydronephrosis severity or the composition of the non-disease group. However, it is reasonable to suppose that the study population may represent sicker patients with more severe forms of urologic disease compared to community-dwelling adults experiencing lower urinary tract symptoms.

The broader lesson for readers of diagnostic studies is this: whenever a diagnostic tool is evaluated in a sample of patients with unusually high disease prevalence, ask who was included in both the disease and non-disease groups. High prevalence is often a signal that the population was pre-selected by referral, severity, or clinical suspicion—and that selection can simultaneously inflate sensitivity by enriching for obvious disease and distort specificity by enriching the non-disease group with look-alike pathology. Reported test characteristics are only as generalizable as the spectrum of disease they were measured against.

Conclusions

The model described in this study raises an intriguing possibility: expanding renal POCUS by letting bedside nurses use handheld ultrasound devices, with AI assistance, to evaluate for hydronephrosis. The authors conclude that the tool is feasible for nurse-led triage. These results are encouraging but should be interpreted with caution. First, the study was small. With only 15 patients with hydronephrosis and 17 without, the estimates are imprecise: the 95% confidence interval for sensitivity spans 70% to 99%, and for specificity 59% to 94%. Second, the prevalence of hydronephrosis in the study population was far higher than even our most generous estimates for community-dwelling adults, so the reported predictive values will not carry over to the settings where the tool is meant to be used. Third, several features of the study may have made the task easier than it will be in practice. Patients were urology inpatients with a clinical indication for ultrasound, likely enriching the sample for more advanced disease. Additionally, only the kidney with the clearest acoustic window was analyzed. Future validation should report the severity of hydronephrosis, even if the tool is not trained to grade it and describe the composition of the non-disease group. It should be conducted in a population whose prevalence and spectrum of disease approximate those of the intended setting, with bilateral scanning, and a sample large enough to estimate test characteristics more precisely. AI evaluation of nurse-obtained images for prostatic enlargement is even less certain. Fifteen percent of prostate scans were non-diagnostic, and with only seven enlarged prostates, sensitivity could plausibly be anywhere from about 36% to 92%. For now, this tool is a promising proof of concept, but it is not ready to replace clinician judgment or formal imaging in real-world triage.

Disclosure Statement

The authors declare that they have no conflict of interest.

Artificial Intelligence Statement

The authors used Claude (Sonnet 5.5) large language model for copy-editing only. The authors take full responsibility for the publication.

References

1. Matsumoto M, Suzuki M, Karube M, Gotoh H, Sato M, Kobayashi T, Masuda M, Wakita S, Kawamoto A, Miura Y, Tamai N, Sanada H. Nurse-validated handheld POCUS artificial intelligence (AI) for automated prostate volume estimation and hydronephrosis detection. POCUS J. 2026;11(2):60-70. doi:10.24908/pocusj.v11i02.20395

2. Newman TB, Kohn MA. Evidence-based diagnosis: an introduction to clinical epidemiology. 2nd ed. Cambridge, UK: Cambridge University Press, 2020.

3. Darrad M, Gupta A, Rukin N. Hydronephrosis. In Blandy’s Urology, Aboumarzouk OM (Ed.). 2019. doi:10.1002/9781118863343.ch10

4. Sinha S, Matai L. Is isolated bladder outlet obstruction associated with hydronephrosis? A database analysis. Neurourol Urodyn. 2020;39(8):2361-2367. doi:10.1002/nau.24495

5. Kohn MA, Carpenter CR, Newman TB. Understanding the direction of bias in studies of diagnostic test accuracy. Acad Emerg Med. 2013;20(11):1194-1206. doi:10.1111/acem.12255

Posted in .

Leave a Reply

Your email address will not be published. Required fields are marked *