The cross-sectional study recruited adults aged 60 or older who received subcutaneous insulin from endocrinology outpatient clinics and inpatient wards between January 2025 and January 2026. Of 1,000 people screened, 182 were excluded because visual inspection or palpation had already identified an injection-site abnormality. The remaining 818 participants underwent standardized visual and palpation examinations followed by high-frequency ultrasound. Ultrasound identified 129 cases of subclinical lipohypertrophy and 689 negative examinations, producing the observed detection proportion of 15.8% (95% confidence interval, 13.4%–18.4%). Mean age was 73.17 years with a standard deviation of 6.46, 60.8% of participants were women, and mean body mass index was 24.75 kg/m² with a standard deviation of 3.18.

Investigators developed two ridge logistic models. A nine-predictor core model included sex, age, body mass index, years of insulin injections, daily injection frequency, needle-reuse count, regular site rotation, self-reported injection area, and immediate injection after removing insulin from refrigeration. An extended model added glycated hemoglobin and a questionnaire-derived measure of weekly glucose-variability frequency. Predictors were collected before ultrasound outcomes were ascertained, while the personnel collecting them remained unaware of ultrasound results and the ultrasound readers had no access to model predictions. In the extended model, the largest absolute standardized coefficients were associated with irregular site rotation, glucose-variability frequency, needle reuse, immediate injection after refrigeration, and age. These coefficients describe predictive associations within the fitted model and do not establish that any individual behavior caused a lesion.

Ultrasound positivity required a reproducible lesion meeting at least three of four specified features: a relatively well-defined subcutaneous nodule with a hypoechoic or hyperechoic halo, distortion of surrounding connective-tissue architecture, absence of internal vascular signals and a capsule, and a heterogeneous echotexture relative to surrounding tissue. Two experienced sonographers interpreted the examinations independently, and a third senior sonographer adjudicated disagreements. A lesion meeting the imaging criteria but not identified as definite lipohypertrophy during visual inspection or palpation was classified as subclinical. This entry population is important because the 15.8% figure applies specifically to patients with negative routine examinations and should not be interpreted as a population-wide prevalence estimate.

Internal validation used 2,000 paired ordinary bootstrap resamples, with design-matrix construction, inner folds, penalty selection, model fitting, prediction, calibration, and scoring repeated within each resample. Repeated nested cross-fitting, using stratified 10-fold outer splits repeated 20 times, generated patient-level out-of-fold probabilities. After bootstrap optimism correction, the core model had a receiver operating characteristic area under the curve of 0.758 (95% interval, 0.720–0.808) and a Brier score of 0.117 (0.100–0.130). The extended model had a corresponding area under the curve of 0.774 (0.741–0.823), a precision-recall area under the curve of 0.423 (0.353–0.527), and a Brier score of 0.113 (0.097–0.126). For the extended model, mean predicted probability was 0.158, matching the observed positive proportion, the calibration intercept was 0.004 (−0.005 to 0.014), and mean absolute calibration error was 0.013 (0.013–0.037).

The incremental contribution of glycated hemoglobin and glucose-variability frequency remained uncertain. Compared with the core model, the extended model’s paired differences were 0.015 in receiver operating characteristic area under the curve, 0.037 in precision-recall area under the curve, and −0.003 in Brier score; the respective 95% intervals all included zero. The researchers therefore identified the core model as a candidate for evaluation as a lower-burden option when metabolic information is unavailable, but the study did not establish clinical equivalence or demonstrate that either model should be preferred. Prespecified sensitivity analyses examined alternative penalty selection, factor coding of injection frequency, unpenalized logistic regression, natural splines, and robust scaling. Across these analyses, cross-fitted areas under the curve for the extended model ranged from 0.764 to 0.771, and Brier scores ranged from 0.114 to 0.117.

Out-of-bag validation assessed thresholds selected to target 90% and 95% sensitivity. The development threshold for t90 was 0.084466. At this operating point, mean sensitivity was 89.7% (95% interval, 77.4%–98.2%), specificity was 38.1% (22.2%–53.3%), positive predictive value was 21.5% (16.5%–26.5%), and negative predictive value was 95.4% (91.6%–99.0%). Per 1,000 patients, the model referred 663 people for ultrasound, avoided 337 examinations, and missed 16 lesions. The relatively low specificity means that many people without ultrasound-positive lesions would also be referred, while the positive predictive value reflects the 15.8% detection proportion among the selected patients entering the model.

The more conservative t95 threshold was 0.068444. Internal out-of-bag validation produced mean sensitivity of 94.4% (83.7%–100.0%), specificity of 25.1% (8.5%–42.1%), positive predictive value of 19.2% (14.9%–23.8%), and negative predictive value of 96.5% (92.2%–100.0%). Per 1,000 patients, 780 would be referred, 220 examinations would be avoided, and nine cases would be missed. Moving from t90 to t95 therefore required about 117 additional examinations for 7.4 fewer missed cases per 1,000. Decision-curve point estimates favored selective referral over either examining everyone or referring no one across threshold probabilities of 0.035–0.30, with net-benefit estimates of 0.116, 0.087, and 0.046 at threshold probabilities of 0.05, 0.10, and 0.20. Whether that trade-off is useful depends on local imaging capacity and the consequences of missing a lesion.

Several features limit the conclusions. This was a cross-sectional, single-center study in China, so it estimated the probability of current ultrasound-detected lesions rather than future incidence, treatment response, or patient benefit. The 129 positive outcomes limited precision for small performance differences and threshold estimates. Self-reported needle reuse, site rotation, injection area, and injection after refrigeration may have been affected by recall and social-desirability biases, with an uncertain effect on model weights and calibration. The glucose-variability measure depended partly on recalled monitoring history and individualized glycemic targets rather than being derived solely from continuous glucose monitoring. Different equipment, scanning protocols, and readers could affect calibration in other settings. The researchers did not assess performance across sociodemographic groups, frailty levels, cognitive function, or caregiving circumstances. Referral classifications were also unstable for 17.6% of participants near t90 and 12.0% near t95, although the median within-participant standard deviation of repeated out-of-fold probabilities was 0.008.

The findings do not establish that the model is ready for clinical software deployment or widespread protocol adoption. The researchers call for temporal and multicenter geographic external validation using standardized ultrasound procedures, with evaluation of calibration, fixed operating points, and subgroup performance. Prospective studies would then need to compare routine assessment with model-assisted referral and measure effects on missed lesions, imaging use, and patient outcomes. Until such evaluation is completed, the study’s central result is that an internally validated model can stratify ultrasound-detected subclinical lipohypertrophy among older insulin-treated adults with negative routine examinations, while its two operating points quantify the numerical trade-off between selective imaging and missed lesions in the study population.

Subject of Research: Internal validation of a diagnostic triage model for ultrasound-detected subclinical insulin injection–related lipohypertrophy in older adults with diabetes

Article Title: Diagnostic triage for ultrasound-detected subclinical insulin injection–related lipohypertrophy in older adults with diabetes: development and internal validation

Article References: Yu, S., Wang, Y., Wang, W., Liu, Z., & Wu, Q. (2026). Diagnostic triage for ultrasound-detected subclinical insulin injection–related lipohypertrophy in older adults with diabetes: development and internal validation. Frontiers in Endocrinology, 17, 1997042. https://doi.org/10.3389/fendo.2026.1997042

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1997042

Keywords: insulin lipohypertrophy, subclinical lesions, diagnostic triage, ultrasound, older adults, diabetes, prediction model, ridge regression, injection technique, internal validation, Diagnostic, triage