CoChat
Sign Up
Cited by 2Diabetes Management and ResearchIntravenous Infusion Technology and SafetyHealthcare Technology and Patient Monitoring

Development of a Real-time Force-based Algorithm for Infusion Failure Detection

Open access
Blanco, Luis E., Wilcox, John H., Hughes, Michael S., Lal, Rayhan A.·2024·Journal of Diabetes Science and Technology
DOI 10.1177/19322968241247530PMC PMC11535327
Open this paper in CoChat — free accountOpen PDF

Abstract

Background: Continuous subcutaneous insulin infusion (CSII) is a common treatment option for people with diabetes (PWD), but insulin infusion failures pose a significant challenge, leading to hyperglycemia, diabetes burnout, and increased hospitalizations. Current CSII pumps’ occlusion alarm systems are limited in detecting infusion failures; therefore, a more effective detection method is needed. Methods: We conducted five preclinical animal studies to collect data on infusion failures, utilizing both insulin and non-insulin boluses. Data were captured using in-line pressure and flow rate sensors, with additional force data from CSII pumps’ onboard sensors in one study. A novel classifier model was developed using this dataset, aimed at detecting different types of infusion failures through direct utilization of force sensor data. Performance was compared against various occlusion alarm thresholds from commercially available CSII pumps. Results: The testing dataset included 251 boluses. The Bagging classifier model showed the highest performance metrics among the models tested, exhibiting high accuracy (96%), sensitivity (94%), and specificity (98%), with lower false-positive and false-negative rate compared with traditional occlusion alarm pressure thresholds. Conclusions: Our study developed a novel non-threshold classifier that outperforms current occlusion alarm systems in CSII pumps in detecting infusion failures. This advancement has the potential to reduce the risk of hyperglycemia and hospitalizations due to undetected infusion failures, offering a more reliable and effective CSII therapy for PWD. Further studies involving human participants are recommended to validate these findings and assess the classifier’s performance in a real-world setting.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper develops a non-threshold machine-learning classifier using CSII pump onboard force sensor data to detect infusion failures, trained and tested on preclinical swine bolus data. The Bagging classifier outperformed commercial occlusion alarm thresholds in retrospective testing, but human validation is still needed.

Key findings

  • In the 251-bolus testing set (169 functional, 82 dysfunctional), the Bagging classifier with 29 features achieved 96% accuracy, 94% sensitivity, 97% specificity, 3% FPR, 6% FNR, and MCC 0.909; all five final models had accuracy ≥90%, but only BAG had MCC ≥0.9.
  • Compared with force-converted commercial occlusion thresholds, BAG had fewer false positives and false negatives; for example, the 13.15 PSI/10,885 mN threshold had 60% accuracy, 30% sensitivity, and 74% specificity, while the 35 PSI/21,783 mN threshold had 67% accuracy, 0% sensitivity, and 100% specificity.
  • All classifications were observed within 13 to 120 seconds following bolus completion and depended on bolus size and injection speed.
  • During testing, LR had the highest PR-AUC (0.979) and ROC-AUC (0.988), followed by SVM, BAG, RF, and GBM.

Methods

  • Five preclinical swine studies conducted between November 2020 and August 2023 yielded 866 boluses; 615 (71%) from studies 1–4 were used as the training set with pressure data only, and 251 (29%) from study 5 were used as the testing set with both onboard force and in-line pressure data.
  • Boluses were manually labeled “Functional” or “Dysfunctional” by two trained individuals using flow rate, pressure, and force data, with a 70% volume-delivery cutoff and six testing-set label overwrites.
  • A third-order polynomial pressure-to-force scaling function was fitted from synchronized study 5 data; models tested included LR, SVM, RF, LDA, GNB, GPC, BAG, and GBM, with feature elimination, hyperparameter tuning, SMOTE, 10-fold 3-repeat stratified k-fold cross-validation, and F2-score optimization.

Limitations

  • The paper states reliance on non-human animal data for both training and testing; ideal validation would involve human data and glucose outcomes.
  • The paper states that training included ambulatory insulin infusion set data but testing did not, and that extended basal delivery scenarios were not tested; pressure-to-force conversion meant training used processed rather than raw data, potentially introducing model bias.
  • The paper states the model identifies individual bolus failures but does not yet provide a comprehensive infusion-site health assessment or actionable replacement recommendations; evident from the text, testing came from a single study with 82 dysfunctional boluses, and no human clinical validation was performed.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1177/19322968241247530

Cites

27 of 27 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

10 of 16 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Extended wear infusion sets present a wider timeframe for infusion failure.

    “given the recent introduction of extended wear infusion sets, 32 which present a wider timeframe for infusion failure”
    Quote in the paperSubstantial claim1 of 1 sources checked
  • Premisecitation

    Current occlusion alarm systems cannot detect other types of infusion failures such as leakages, dislodged cannulas, and silent occlusions.

    “However, these detection thresholds are not standardized across manufacturers and cannot detect other types of infusion failures such as leakages, dislodged cannulas, and “silent” (ie, no alert) occlusions.15,16”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [15] source

      The Discussion (page unavailable) says current pumps have no leakage-detection alarms and that most cannula malplacements may go unrecognized for hours. It also says some malplacements are eventually recognized by high-pressure alarms, so an absolute inability to detect them overstates the source. The paper does not address “silent” occlusions. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [16] source

      The Mechanical problems and Hyperglycemia sections identify leaks, displaced catheters, and blocked infusion sets as causes of unexplained high blood glucose; Mechanical problems also says leaks can be difficult to detect. Neither section says that occlusion alarms fail to detect these problems or silent occlusions, and the paper does not establish that alarm thresholds vary across manufacturers. No page numbers are available in our copy. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
  • Premisecitation

    Occlusion detection thresholds are not standardized across manufacturers.

    “However, these detection thresholds are not standardized across manufacturers and cannot detect other types of infusion failures such as leakages, dislodged cannulas, and “silent” (ie, no alert) occlusions.15,16”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [15] source
    • cites [16] source

      The Mechanical problems section discusses clogged infusion sets, leakage, and displaced sets, but does not address occlusion-detection thresholds or whether manufacturers standardize them. Searches for threshold standardization and occlusion-alarm pressure found no relevant passage in our copy; page numbers were not available. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • Premisecitation

    Testing of retrofitted pressure sensors in CSII pumps suggests that the incorporation of more precise fluid monitoring could improve infusion failure detection.

    “Alternatively, the testing of retrofitted pressure sensors in CSII pumps suggests that the incorporation of more precise fluid monitoring could improve the infusion failure detection.15”
    Quote in the paperSubstantial claim1 of 1 sources checked
  • Premisecitation

    Occlusion alarms can take time to trigger depending on the pump's infusion rate, with delays up to 40 hours post-failure.

    “Furthermore, occlusion alarms can take time to trigger depending on the pump's infusion rate, with delays up to 40 hours post-failure.17”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [17] source

      The Results report that occlusion detection took substantially longer at 0.1 U/h than at 1.0 U/h, supporting the infusion-rate dependence. But “up to 40 hours” understates the observed delays: the longest median was 41 hours 21 minutes, and an individual alarm took 65 hours 40 minutes. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
  • Classified as probably not a substantial claim. 17
  • May combine two claims. 25
  • The quote only partly states this claim. 2

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    During training, model performance was evaluated using 10-fold, 3-repeats stratified k-fold cross-validation and F2-score.

    “During training, model performance was evaluated using 10-fold, 3-repeats stratified k-fold cross-validation and F2-score (ie, instead of F1-score) as it penalizes false negatives.24,25”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [24] source

      The Methods section (Model evaluation metrics; page unavailable) reports stratified tenfold cross-validation on the training set. It does not report three repeats or use of the F2-score; instead, it names accuracy, AUC, recall, and precision as its evaluation metrics. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites [25] source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Infusion failures are the leading cause of hospitalizations for diabetes-related ketoacidosis among people with diabetes who use CSII therapy, in some cases resulting in fatal outcomes.

    “Notably, infusion failures are the leading cause of hospitalizations for diabetes-related ketoacidosis (DKA) hospitalization among PWD who use CSII therapy, in some cases resulting in fatal outcomes.8,9”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [8] source

      The Results (Etiology of DKA; page not provided) report that pump/tubing defects were associated with 55% of DKA cases among the study’s CSII users, making them the leading cause in that hospitalized group. The Discussion describes DKA as potentially fatal but does not link any fatal outcome specifically to an infusion failure. The claim also generalizes the finding beyond the adults with type 1 diabetes studied. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [9] source

  • May combine two claims.

    Findingown data

    During tuning, all models showcased strong PR-AUC and ROC-AUC which speaks to the model's robustness and ability to generalize past the training set.

    “During tuning, all models showcased strong PR-AUC and ROC-AUC which speaks to the model's robustness and ability to generalize past the training set (Figure 5).24”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [24] source

      The Results sections on model performance and feature selection report ROC-AUC values and use cross-validation and external validation to assess robustness and generalization (page not provided). They do not report PR-AUC, and the decision tree performed poorly on external validation, so the claim that all models had strong results on both measures is not supported. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingnone given

    The model provides immediate detection of failures without requiring glucose or other external input.

    “This model provides immediate detection of failures without requiring glucose or other external input”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    PR AUC shows minimal difference between models, 0.013 difference between best (GBM) and worst model (RF).

    “PR AUC highlights minimal difference between models, 0.013 difference between best (GBM) and worst model (RF).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    All models showed accuracies ≥0.9, but only BAG had an MCC ≥0.9.

    “Overall, all models showed accuracies ≥0.9, but only BAG had an MCC ≥0.9.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • May combine two claims.

    Findingown data

    A total of 251 (29%) boluses (study 5) were used for the testing set and had data from both the study pump's onboard force sensor and the in-line pressure sensor.

    “A total of 251 (29%) boluses (study 5) were used for the testing set and had data from both the study pump's onboard force sensor and the in-line pressure sensor.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    BAG (29 features) had the highest accuracy (96%), MCC (0.909), sensitivity (94%), specificity (97%), FPR (3.0%), and FNR (6.0%) among the five trained models.

    “Of the five, BAG (29 features) had the highest accuracy (96%), MCC (0.909), sensitivity (94%), specificity (97%), FPR (3.0%), and FNR (6.0%).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • May combine two claims.

    Findingown data

    Logistic regression, SVM, RF, BAG, and GBM showed similar performance with Recall > 0.86, Precision > 0.80, and F2-score > 0.86.

    “Logistic regression, SVM, RF, BAG, and GBM showed similar performance with Recall > 0.86, Precision > 0.80, and F2-score > 0.86 so LDA, GNB, and GPC were not considered for further testing (Figure 3, Table 2).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The other 5 models have Recall > 0.86 and Precision > 0.80.

    “Here, we see that LDA, GNB, and GPC are not clustered with the other 5 models (Recall > 0.86, Precision > 0.80) and can be eliminated going forward.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Of the 251 values included in the testing dataset, 169 (67.3%) were labeled as Functional and 82 (32.7%) were Dysfunctional.

    “Of the 251 values included in the testing dataset, 169 (67.3%) were labeled as “Functional” and 82 (32.7%) were “Dysfunctional.””
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • May combine two claims.

    Findingown data

    Of the models, BAG yielded the highest performance; however, even LR (the least complex model) performed nearly as well.

    “Of the models, BAG yielded the highest performance; however, even LR (the least complex model) performed nearly as well.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • May combine two claims.

    Findingnone given

    All classifications were observed within 13 to 120 seconds following bolus completion and were dependent on bolus size and injection speed.

    “All classifications were observed within 13 to 120 seconds following bolus completion and were dependent on bolus size and injection speed.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    All models showcased strong PR-AUC and ROC-AUC scores during testing with LR having the highest in both instances with 0.979 and 0.988, respectively, followed by SVM, BAG, RF, and GBM.

    “All models showcased strong PR-AUC and ROC-AUC scores during testing with LR having the highest in both instances with 0.979 and 0.988, respectively, followed by SVM, BAG, RF, and GBM (Figure 5).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • May combine two claims.

    Findingown data

    A total of 615 (71%) boluses (studies 1-4) were used as the training set and only contained pressure data.

    “A total of 615 (71%) boluses (studies 1-4) were used as the training set and only contained pressure data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The identified optimal feature subsets included 33, 27, 98, 44, and 29 features for LR, SVM, RF, BAG, and GBM, respectively, achieving F2-scores of 0.930, 0.923, 0.940, 0.934, and 0.932.

    “The identified optimal subsets included 33, 27, 98, 44, and 29 features for LR, SVM, RF, BAG, and GBM, respectively, achieving F2-scores of 0.930, 0.923, 0.940, 0.934, and 0.932 (Figure 4).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Overwriting of %VD occurred in six instances in the testing set.

    “Overwriting of %VD occurred in six instances in the testing set.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    The output of the pressure-to-force preprocessing step was a third-order polynomial function used to transform the training set's pressure data to force.

    “The output of this preprocessing step was a thirdorder polynomial function that we could use to transform the training set's pressure data to force (Supplement 3).”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    The BAG model achieved higher accuracy (96%) than the force-based occlusion thresholds (60-68%) on the testing dataset.

    “| BAG | 29 | 77 | 164 | 5 | 5 | 96% | 94% | 97% | 3% | 6% | 0.909 |”
    • Also stated in: § Conclusion · plain
      “| 30 d | 19 084 | 4 | 167 | 2 | 78 | 68% | 5% | 99% | 1% | 95% | 0.113 |”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The testing dataset is representative of the broader population of insulin pump users.

  • The quote only partly states this claim.

    Findingown data

    The BAG model achieved higher sensitivity (94%) than the force-based occlusion thresholds (0-30%) on the testing dataset.

    “| BAG | 29 | 77 | 164 | 5 | 5 | 96% | 94% | 97% | 3% | 6% | 0.909 |”
    • Also stated in: § Conclusion · plain
      “| 13.15 a | 10 885 | 25 | 125 | 44 | 57 | 60% | 30% | 74% | 26% | 70% | 0.047 |”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The testing dataset is representative of the broader population of insulin pump users.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    All study boluses were analyzed simultaneously by two trained individuals and manually assigned a binary label: Functional or Dysfunctional.

    “All study boluses were analyzed simultaneously by two trained individuals and manually assigned a binary label: “Functional” or “Dysfunctional.””
    Quote in the paperSubstantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The model is a robust tool that could be applied across a variety of system configurations and subcutaneous infusion set types.

    “making it a robust tool that could be applied across a variety of system configurations and subcutaneous infusion set types.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The testing dataset is representative of the broader population of insulin pump users.

  • May combine two claims.

    Premisenone given

    CGM-based models require an elevation of glucose levels over time, which may produce false alarms if hyperglycemia develops that is not related to an infusion failure.

    “These models also require an elevation of glucose levels over time, which may produce false alarms if hyperglycemia develops that is not related to an infusion failure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Algorithms designed to detect infusion failure using continuous glucose monitoring and historical insulin infusion data have been developed but not implemented commercially.

    “Algorithms designed to detect infusion failure using continuous glucose monitoring (CGM) and historical insulin infusion data have been developed but not implemented commercially.[19][20][21]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19][20][21]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    This potential universal compatibility could allow broad adoption, offering all users, regardless of their specific devices, the opportunity to enhance the safety and efficacy of their insulin therapy.

    “This potential universal compatibility could allow board adoption, offering all users, regardless of their specific devices, the opportunity to enhance the safety and efficacy of their insulin therapy.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Onboard force sensors across different CSII pump systems and infusion set types provide comparable data suitable for the algorithm.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Standard threshold methods are limited by an inverse trade-off between increasing false positive rate or false negative rate as the threshold is increased or decreased, respectively.

    “Standard threshold methods are limited by an inverse trade-off between either increasing FPR or FNR as the threshold is increased or decreased, respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Each study received prior approval from either the Medical Education & Research Institute or University of Tennessee Health Science Center Institutional Animal Care and Use Committee.

    “Each study received prior approval from either the Medical Education & Research Institute (MERI) or University of Tennessee Health Science Center (UTHSC) Institutional Animal Care and Use Committee (Memphis, Tennessee).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Because the algorithm operates based on data from onboard force sensors, its implementation could be possible across a variety of systems and infusion set types.

    “In addition, because the algorithm operates based on data from onboard force sensors, its implementation could be possible across a variety of systems and infusion set types.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Onboard force sensors across different CSII pump systems and infusion set types provide comparable data suitable for the algorithm.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The testing included small-volume boluses, but did not test more extended basal delivery scenarios.

    “In addition, our testing included smallvolume boluses, but did not test more extended basal delivery scenarios (this type of data was collected during study 4 and will be examined in later work).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Site leakage was not separately identified or quantified.

    “Site leakage was not separately identified or quantified.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    The conversion of pressure to force means training itself was conducted on processed rather than raw data, potentially introducing model bias.

    “Despite effective dataset training, as evidenced by a strong MCC, the conversion means training itself was conducted on processed rather than raw data, potentially introducing model bias.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The ideal validation scenario would involve testing the classifier on human data.

    “While swine models offer a reasonable approximation of human tissue and the use of imaging data allows for confirmation of bolus infusion, the ideal validation scenario would involve testing the classifier on human data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    This capability could empower individuals to address issues with their infusion sites as soon as they occur, thereby reducing the risk of hyperglycemia and preventing more severe complications, such as DKA.

    “This capability could empower individuals to address issues with their infusion sites as soon as they occur, thereby reducing the risk of hyperglycemia and preventing more severe complications, such as DKA.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Swine models offer a reasonable approximation of human tissue.

  • May combine two claims.

    Interpretationown data

    Compared to standard commercial pump occlusion detection thresholds, the models achieve higher accuracy, sensitivity, specificity, and MCC with fewer false positives and false negatives.

    “Compared to standard commercial pump occlusion detection thresholds, we achieve higher accuracy, sensitivity, specificity, and MCC with fewer FPs and FNs.”
    • Also stated in: § Abstract · plain
      “Our study developed a novel non-threshold classifier that outperforms current occlusion alarm systems in CSII pumps in detecting infusion failures.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Extended wear infusion sets are prime candidates for active monitoring.

    “making them prime candidates for the benefits of an active monitoring system.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Insulin infusion failures in CSII are associated with reduced insulin delivery, prolonged hyperglycemia, and increased likelihoods of diabetes burnout and pump discontinuation.

    “Infusion failures are associated with reduced insulin delivery, prolonged hyperglycemia, and increased likelihoods of diabetes burnout and pump discontinuation.[3][4][5]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3][4][5]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    This advancement has the potential to reduce the risk of hyperglycemia and hospitalizations due to undetected infusion failures, offering a more reliable and effective CSII therapy for people with diabetes.

    “This advancement has the potential to reduce the risk of hyperglycemia and hospitalizations due to undetected infusion failures, offering a more reliable and effective CSII therapy for PWD.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Swine models offer a reasonable approximation of human tissue.
37 of 37

  • The paper relies on this without support.

    Swine models offer a reasonable approximation of human tissue.

    “While swine models offer a reasonable approximation of human tissue and the use of imaging data allows for confirmation of bolus infusion, the ideal validation scenario would involve testing the classifier on human data.”

  • The paper relies on this without support.

    The manually assigned binary labels based on fluid dynamics data accurately represent functional versus dysfunctional boluses.

    “All study boluses were analyzed simultaneously by two trained individuals and manually assigned a binary label: “Functional” or “Dysfunctional.””

  • The paper relies on this without support.

    Onboard force sensors across different CSII pump systems and infusion set types provide comparable data suitable for the algorithm.

    “In addition, because the algorithm operates based on data from onboard force sensors, its implementation could be possible across a variety of systems and infusion set types.”

  • The paper relies on this without support.

    The testing dataset is representative of the broader population of insulin pump users.

    “Performance of the Models and Occlusion Alarm Thresholds on the Testing Dataset (Total N = 251, 169 Functional, 82 Dysfunctional).”
4 of 4

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • BAG does not have fewer false positives or higher specificity than every threshold

    Major

    The discussion claims higher specificity and fewer false positives than commercial thresholds. But Table 3 gives BAG 97% specificity and 5 false positives, versus 100% specificity and 0 false positives at the 35 PSI threshold.

  • Abstract overstates BAG specificity

    Minor

    The abstract reports 98% specificity for BAG, whereas the Results report 97%. Table 3 also reports 97% (164 true negatives among 169 functional boluses).

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Blanco, L. E., Wilcox, J. H., Hughes, M. S., & Lal, R. A.
Affiliations
  • Diatech Diabetes, Inc., Memphis, TN, USA
  • Stanford University (US)
ORCID
Venue
Journal of Diabetes Science and Technology
Publisher
SAGE Publications
Published
2024-04-23
Volume
18
Issue
6
Pages
1313-1323
ISSN
1932-2968 (ISSN-L)
DOI
10.1177/19322968241247530
PMC
PMC11535327
OpenAlex
W4395078941
Semantic Scholar
b831013f71cd21039dd4064ec2c78f21337100be
Cited by
2
References
27
Citations by year
2025: 1 · 2024: 1
Type
conference abstract
Language
en
Open access
Licence
journals.sagepub.com/page/policies/text-and-data-mining-license (tdm)
Locations
  • Journal of Diabetes Science and Technology (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion)
Funders
  • The Consortium for Technology & Innovation in Pediatrics (P50FD007963)
  • National Institute of Diabetes and Digestive and Kidney Diseases (1K23DK122017, 1K23DK138267, 1R43DK130036-01, P30 DK116074)
  • Ypsomed AG
  • JDRF (3-IND-2023-1444-I-X)
Topics
  • Health Sciences › Medicine › Endocrinology, Diabetes and Metabolism › Diabetes Management and Research
  • Physical Sciences › Engineering › Biomedical Engineering › Intravenous Infusion Technology and Safety
  • Health Sciences › Medicine › Surgery › Healthcare Technology and Patient Monitoring
Keywords
insulin pump therapy, occlusion detection, insulin pump, bagging classifier, diabetes mellitus, hyperglycemia prevention, real-time classification
MeSH terms
Algorithms, Animals, Equipment Failure, Humans, Hypoglycemic Agents, Insulin, Insulin Infusion Systems, Male, Clinical Alarms

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review