Skip to main content
Tessa NewsLink
Paste a health news link, or browse

Source study found

Story checked

AI tool designed and integrated into hospital workflows to classify skin lesions (opens in a new tab)

medicalxpress.com · 2026-10-05

Short answerEvidenceSource

Short answer

Not supported

Not supported.

One key claim is not backed by the study. 6 other points were not covered by the paper.

  • 1 overstated
  • 1 not supported
  • 6 not covered

Checked against the study summary. The full text wasn't available, so some details couldn't be settled either way.

Share this check

Follow the evidence trail
1
2

NewsLink checks it

Not supported

Two claims go beyond the study. One overstates it and one isn't supported at all. Six claims the study doesn't address.

  • 1 overstated
  • 1 not supported
  • 6 not covered
Open claim evidence
3
Then inspect each claim

Evidence layer

Claim by claim

Each claim gets a verdict. Expand it to see the evidence directly below.

8 claims in this story

Showing all 8 claimsChoose a verdict to focus the list.

Claim 1 of 8Not supportedThe article says the system achieved 86% overall accuracy, 88% sensitivity for melanoma detection and 92% sensitivity for basal cell carcinoma detection.View evidenceHide evidence

As stated86% overall accuracy; 88% melanoma sensitivity; 92% basal cell carcinoma sensitivity

Why this verdict

The 86% overall accuracy aligns with the internal cross-validation accuracy reported in the abstract profile. However, the stated 88% melanoma sensitivity and 92% basal-cell-carcinoma sensitivity are not supported by the supplied profile; the external evaluation reports melanoma sensitivity 0.60 and BCC sensitivity 0.72. The story also does not distinguish internal cross-validation from external testing.

1

Study evidence

Stratified 5-fold cross-validation on ISIC 2019 BCN_20000 and MSK subsets yielded aggregate performance reported in the abstract: accuracy 0.86, macro F1-score 0.85, and AUC values above 0.97.accuracy 0.86; macro F1 0.85; AUC > 0.97

“Internal robustness was assessed through stratified 5-fold cross-validation”
2

Study evidence

On the independent HAM10000 external test set, the model achieved accuracy 0.84, balanced accuracy 0.65, with sensitivities of 0.60 for melanoma and 0.72 for basal cell carcinoma.accuracy 0.84; balanced accuracy 0.65; sensitivity (melanoma) 0.60; sensitivity (BCC) 0.72

“external generalization was evaluated on the independent HAM10000 dataset.”
Claim 2 of 8OverstatedThe technology supports health care staff in evaluating lesions while integrating into the hospital's existing IT infrastructure.View evidenceHide evidence

Why this verdict

The profile supports integration with hospital IT workflows and a motivation to support skin-lesion assessment. But the story's unhedged causal framing that the technology supports health care staff in evaluating lesions goes beyond the abstract evidence, which reports technical/model performance and interoperability but says clinical impact on decision-making, workflow, or patient outcomes was not assessed.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'

“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
Claim 3 of 8Not coveredThe University of Alicante and Sant Joan d'Alacant University Hospital collaborated to launch MEL-IA, an artificial intelligence system designed to automate the classification of skin lesions.View evidenceHide evidence

Why this verdict

The abstract-level profile supports that MEL-IA is an AI/multimodal system for automated skin-lesion classification and reports hospital deployment. However, the specific institutional collaboration and the framing that the institutions 'launched' the system are not established in the supplied abstract profile.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'

“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
Claim 4 of 8Not coveredThe system includes a mobile app to capture lesion images and clinical data, an AI model to classify lesions, and secure integration with hospital systems.View evidenceHide evidence

Why this verdict

The AI model and hospital-system integration are supported at abstract level: the system combines dermatoscopic images with structured clinical metadata and integrates with HL7, DICOM, PACS, and HIS/RIS workflows. The supplied profile does not verify the mobile app component or the claim of secure integration.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'

“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
Claim 5 of 8Not coveredUnlike tools that only separate malignant from benign lesions, MEL-IA provides differential categorization across five lesion types: melanoma, nevus, basal cell carcinoma, actinic keratosis and benign keratosis.View evidenceHide evidence

As statedfive major skin lesion types

Why this verdict

The profile supports multiclass lesion classification and mentions melanoma and basal cell carcinoma metrics, but the abstract profile does not provide the exact five-category list, nor does it support the comparison with tools that only separate malignant from benign lesions.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

On the independent HAM10000 external test set, the model achieved accuracy 0.84, balanced accuracy 0.65, with sensitivities of 0.60 for melanoma and 0.72 for basal cell carcinoma.accuracy 0.84; balanced accuracy 0.65; sensitivity (melanoma) 0.60; sensitivity (BCC) 0.72

“external generalization was evaluated on the independent HAM10000 dataset.”
Claim 6 of 8Not coveredResearchers trained and validated the model using over 15,000 dermatoscopic images plus clinical patient data such as age, sex and lesion location.View evidenceHide evidence

As statedover 15,000 dermatoscopic images

Why this verdict

The profile supports training on ISIC 2019 BCN_20000 and MSK subsets, use of dermatoscopic images plus structured clinical metadata, internal cross-validation, and external testing. The abstract profile does not verify the specific 'over 15,000' image count or the particular metadata fields of age, sex, and lesion location.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

Stratified 5-fold cross-validation on ISIC 2019 BCN_20000 and MSK subsets yielded aggregate performance reported in the abstract: accuracy 0.86, macro F1-score 0.85, and AUC values above 0.97.accuracy 0.86; macro F1 0.85; AUC > 0.97

“Internal robustness was assessed through stratified 5-fold cross-validation”
Claim 7 of 8Not coveredThe system was deployed at Sant Joan d'Alacant University Hospital to evaluate performance in a live health care environment, with 980 dermatological studies processed in under one second response times.View evidenceHide evidence

As stated980 dermatological studies; under one second

Why this verdict

The profile supports deployment in a real hospital environment and near-real-time processing through HL7, DICOM, PACS, and HIS/RIS integration. But the supplied abstract profile does not verify the named hospital, the number of 980 dermatological studies, or response times under one second; it only reports >99% successful study integration and 'near real-time' processing.

1

Study evidence

In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'

“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
Claim 8 of 8Not coveredThe researchers emphasized that MEL-IA is a clinical decision-support tool rather than an autonomous diagnostic system, and said next steps include prospective studies with health care professionals, smartphone image capture, and expansion to more lesion types.View evidenceHide evidence

Why this verdict

The profile supports the need for further clinical validation to assess impact on diagnostic decision-making and patient outcomes. It does not verify the full story claim that researchers emphasized MEL-IA is not autonomous, nor the specific next steps involving health care professionals, smartphone image capture, and expansion to more lesion types.

1

Study evidence

An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.

“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
2

Study evidence

In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'

“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
Then look for missing context

Context layer

What the story left out

Important study details the story did not include.

  • The model was trained on ISIC 2019 BCN_20000 and MSK subsets and benchmarked against an image-only baseline, with the abstract stating the multimodal model outperformed the image-only baseline across all classes.

    The story mentions training/validation data in general terms but does not report the named datasets or the image-only baseline comparison. Its comparison to binary malignant/benign tools is not the same as the paper's image-only baseline.

    From secondary_data, supervised deep learning

  • External generalization was evaluated on HAM10000, with accuracy 0.84, balanced accuracy 0.65, melanoma sensitivity 0.60, and basal cell carcinoma sensitivity 0.72.

    The story does not report the external HAM10000 evaluation or its lower balanced accuracy and sensitivities. Instead, it gives higher melanoma and BCC sensitivities not supported by the supplied abstract profile.

    From External validation on HAM10000

  • External performance showed lower balanced accuracy than internal metrics, indicating potential domain differences and variable per-class performance.

    This interpretation-changing limitation is not reflected in the story. The story reports favorable headline performance numbers without noting the lower external balanced accuracy or lower external sensitivities for melanoma and BCC.

    From External validation on HAM10000

  • The abstract lacks key methodological and uncertainty details, including confidence intervals, per-fold variability, sample sizes for external evaluation, class balance, preprocessing, and hyperparameter/model-selection details.

    The story does not communicate these abstract-level limitations, which affect how strongly readers should interpret the reported model-performance metrics.

    From secondary_data, supervised deep learning; Stratified 5-fold cross-validation; External validation on HAM10000

  • The deployment evaluation did not assess clinical workflow impact, user acceptance, diagnostic decision-making, or patient-level outcomes, and the abstract does not define the operational metrics in detail.

    The story mentions decision support and future studies but does not clearly state that the reported deployment was a technical interoperability/operational evaluation rather than evidence of clinical benefit or user/workflow impact.

    From deployment/integration validation

4 things the story did carry across
  • MEL-IA is an EfficientNet-B4–based multimodal skin-lesion classifier combining dermatoscopic images with structured clinical metadata.
  • Internal robustness was assessed using stratified 5-fold cross-validation, with global accuracy 0.86, macro F1-score 0.85, and AUC values above 0.97.
  • MEL-IA deployment in a real hospital environment integrated with HL7, DICOM, PACS, and HIS/RIS workflows and reported >99% successful study integration with near-real-time processing.
  • The abstract-level profile states further clinical validation is warranted to assess impact on diagnostic decision-making and patient outcomes.
Then read the study layer

Study layer

Study at a glance

Scan the study first. Expand only the parts you want to inspect.

Pieces of work

4

Evidence read

study summary

Lead result

secondary data

1Lead resultsecondary dataDevelop a multimodal AI model (EfficientNet-B4–based) that combines dermatoscopic images and structured clinical metadata to classify skin lesions, and benchmark it against an image-only baseline.secondary data, supervised deep learningExpand

In plain English

Developed MEL-IA: an EfficientNet-B4–based multimodal skin lesion classifier that fuses dermatoscopic images with structured clinical metadata, trained on ISIC 2019 (BCN_20000) and MSK subsets and compared against an image-only baseline. Internal stratified 5-fold cross-validation reported high performance (accuracy 0.86, macro F1 0.85, AUC > 0.97). External evaluation on HAM10000 showed accuracy 0.84 and balanced accuracy 0.65 (sensitivities: melanoma 0.60, basal cell carcinoma 0.72). The abstract states the multimodal model outperformed the image-only baseline across all classes.

Key findings

  • An EfficientNet-B4–based multimodal model was developed that combines dermatoscopic images with structured clinical metadata and was trained on ISIC 2019 BCN_20000 and MSK subsets.
  • Internal stratified 5-fold cross-validation showed high internal performance: global accuracy 0.86, macro F1-score 0.85, and AUC values above 0.97.accuracy 0.86; macro F1 0.85; AUC > 0.97
“The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets.”
What this piece can’t prove
  • External performance shows lower balanced accuracy (0.65) than internal metrics, indicative of domain differences; the abstract does not detail causes or dataset label harmonization.
  • Authors state that further clinical validation studies are warranted to assess impact on decision-making and patient outcomes.

1 further detail could not be confirmed from the summary.

2secondary dataEvaluate model robustness and generalization via internal cross-validation on ISIC 2019 (BCN_20000 and MSK subsets) and external testing on HAM10000.Stratified 5-fold cross-validationExpand

In plain English

Internal robustness of a multimodal EfficientNet-B4–based skin lesion classifier was evaluated using stratified 5-fold cross-validation on ISIC 2019 (BCN_20000 and MSK subsets); the abstract reports aggregate discrimination metrics (accuracy, macro F1, AUC) from this procedure.

Key findings

  • Stratified 5-fold cross-validation on ISIC 2019 BCN_20000 and MSK subsets yielded aggregate performance reported in the abstract: accuracy 0.86, macro F1-score 0.85, and AUC values above 0.97.accuracy 0.86; macro F1 0.85; AUC > 0.97
“Internal robustness was assessed through stratified 5-fold cross-validation”
What this piece can’t prove
  • The abstract does not describe whether hyperparameter tuning or model selection was performed within the cross-validation framework.

3 further details could not be confirmed from the summary.

3secondary dataEvaluate model robustness and generalization via internal cross-validation on ISIC 2019 (BCN_20000 and MSK subsets) and external testing on HAM10000.External validation on HAM10000Expand

In plain English

External generalization was evaluated on the independent HAM10000 dataset; the model achieved accuracy 0.84, balanced accuracy 0.65, with sensitivity 0.60 for melanoma and 0.72 for basal cell carcinoma, as reported in the abstract.

Key findings

  • On the independent HAM10000 external test set, the model achieved accuracy 0.84, balanced accuracy 0.65, with sensitivities of 0.60 for melanoma and 0.72 for basal cell carcinoma.accuracy 0.84; balanced accuracy 0.65; sensitivity (melanoma) 0.60; sensitivity (BCC) 0.72
“external generalization was evaluated on the independent HAM10000 dataset.”
What this piece can’t prove
  • Abstract provides no sample size or uncertainty estimates (e.g., confidence intervals) for external metrics.
  • Limited per-class reporting (sensitivities for melanoma and BCC only); other classes and metrics not reported.
  • No information in abstract about calibration, thresholding, or handling of class imbalance on the external set.

1 further detail could not be confirmed from the summary.

4otherDemonstrate hospital interoperability and operational performance of the MEL-IA system via deployment/integration with HL7, DICOM, PACS, and HIS/RIS workflows.deployment/integration validationExpand

In plain English

The study reports deployment of the MEL-IA multimodal skin‑lesion classification system in a real hospital environment, integrated with HL7, DICOM, PACS and HIS/RIS workflows. Technical deployment testing measured interoperability and operational performance, reporting >99% successful study integration and near real‑time processing, indicating feasibility of standards-based integration into hospital infrastructures.

Key findings

  • In a real-hospital deployment, MEL-IA integrated with HL7, DICOM, PACS and HIS/RIS workflows and achieved >99% successful study integration with near real‑time processing, indicating technical interoperability and operational viability.>99% successful study integration; processing described as 'near real-time'
“Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems.”
What this piece can’t prove
  • Key operational metrics lack definitions and numeric detail (e.g., exact latency numbers, counts of integrated studies).
  • No reported assessment of clinical workflow impact, user acceptance, or patient-level outcomes within the deployment.

2 further details could not be confirmed from the summary.

Finally, the search trail

Method layer

NewsLink found the paper. Tessa takes you deeper.

NewsLink checks the story. Tessa is where you inspect the paper, authors, evidence, and research context.

Papers considered

The selected paper, plus nearby candidates.

Europe PMC, Crossref, PubMed · 37 candidate papers

Candidate

Artificial Intelligence-Based Pest and Disease Detection Systems in Precision Agriculture

American Journal of Advanced Medical and Surgical Sciences · 2026 · Crossref

Source (opens in a new tab)

And 31 more candidates considered.