AI-Based Automatic Psoriasis Area and Severity Index (APASI) System Effective in Patients

The artificial intelligence (AI)-driven system known as Automatic Psoriasis Area and Severity Index (APASI) is effective in providing objective, rapid, and standardized evaluations of psoriasis severity, new findings suggest.1

These findings resulted from research conducted to develop and validate the APASI assessment of psoriasis severity, authored by such investigators as Gastón Roustán-Gullón, MD, PhD, of Puerta de Hierro Hospital in Madrid. Roustán-Gullón and coauthors highlighted that traditionally, psoriasis is evaluated using the Psoriasis Area and Severity Index (PASI).

PASI still faces several limitations, the investigators noted. A standardized method for PASI scoring was developed, but it did not include a detailed assessment of specific visual signs. Consequently, Roustán-Gullón et al noted the need for more comprehensive AI-based solutions.

“The APASI tool presented here is accessible through the Legit.Health platform,” the investigators wrote.1 “Additionally, to promote transparency and academic collaboration, we provide a detailed technical description of the methodology and the dataset used, enabling reproducibility and further research by the scientific community.”

Study Details and Findings

The investigative team compiled the Legit.Health-PsO-PASI dataset, with a total of 2857 images spanning various disease severities being implemented and annotated by a set of 3 dermatologists. The team developed 2 deep learning modules: 1 for lesion segmentation and the second for visual sign intensity classification (induration, erythema, and desquamation).

Roustán-Gullón and colleagues looked into multiple encoder architectures, including SE_ResNeXt, EfficientNet, Xception, ResNet, Inceptionv4, and SegFormer, evaluating them through the use of a 5-fold cross-validation approach. In their assessment of performance metrics, the investigators looked at accuracy, root mean square error, intersection over union, confusion matrices and Cohen's kappa for assessing inter-observer variability.

Overall, the investigative team concluded that the best-performing model, MiT_b2, attained human-comparable performance. Specifically, the team noted accuracies of 54.3% for induration, 60.6% for erythema, and 61.8% for desquamation. In their assessment of its performance in lesion segmentation, the team found that the Xception model outperformed expert dermatologists, achieving an intersection over union of .752.

APASI represents a transformative AI-driven tool for psoriasis severity assessment, offering objective, rapid, and precise evaluations. Its integration into clinical and research workflows has the potential to standardize scoring, reduce costs, and enhance data quality in both clinical practice and pharmaceutical research.

While Roustán-Gullón et al's findings highlight APASI’s potential to revolutionize psoriasis severity evaluations via automated, objective, and scalable analysis, they did note several limitations. First, while the system eliminates the necessity for a clinician to be present during image collection, the accuracy of APASI scoring depends on obtaining complete and representative images of all affected skin areas.

Additionally, patients with Fitzpatrick skin types V and VI accounted for only 8% of the study's dataset, suggesting limited representation. This underrepresentation may impact the AI tool’s generalizability and performance in those with darker skin tones, as the visual manifestations of psoriasis can vary substantially by skin type.

“While these initial results are promising, they are based solely on atlas data,” they concluded.1 “To address this, future work will focus on clinical validation through in-person PASI scoring and real-world implementation. A structured imaging protocol will be employed to approximate body surface area involvement more accurately, thereby enabling comprehensive APASI scoring in diverse patient populations. This step is crucial for ensuring the model's generalisability, reliability, and clinical applicability.”

References

1. Carthy TM, Dagnino D, Roustán-Gullón G, et al. Artificial Intelligence-Based Quantification to Assess the Automatic Psoriasis Area and Severity Index. JEADV Clinical Practice. https://doi.org/10.1002/jvc2.70143.

2. Huang K, Wu X, Zhao S, et al. Artificial Intelligence-Based Psoriasis Severity Assessment: Real-world Study and Application. J Med Internet Res. 2023 Mar 16;25:e44932. doi: 10.2196/44932. PMID: 36927843; PMCID: PMC10131673.