Modern medicine is supported by increasingly sophisticated algorithms. In diagnostics or screening, statistical models are commonly used to assess the risk of disease development, the severity of its course, and expected treatment outcome. The growing availability of very detailed data and increased interest in personalized medicine are leading to the development of effective but complex machine learning models. For these models to be trusted, their predictions must be understandable to both the physician and the patient, hence the growing interest in the area of Explainable Artificial Intelligence (XAI). In this paper, we present selected methods from the XAI field in the example of models applied to assess lung cancer risk in lung cancer screening through low-dose computed tomography. The use of these techniques provides a better understanding of the similarities and differences between three commonly used models in lung cancer screening, i.e., BACH, PLCO${}_{m2012}$, and LCART. For the presentation of the results, we used data from the Domestic Lung Cancer Database. The XAI techniques help to better understand (1) which variables are most important in which model, (2) how they are transformed into model predictions, and facilitate (3) the explanation of model predictions for a particular screenee.

Gastrointestinal symptoms are common in critically ill COVID-19 patients. There is currently no generally recognized method of assessing gastrointestinal injury in unconscious or sedated intensive care unit (ICU) patients. I-FABP (intestinal fatty acid binding protein) and citrulline have previously been studied as potential biomarkers of enterocyte damage in various gastrointestinal tract diseases, and changes in the levels of these markers may reflect intestinal wall damage in COVID-19. Patients with critical COVID-19, with diagnosed sepsis, or septic shock requiring ICU treatment were included in the study. Blood samples for citrulline and I-FABP were taken daily from day 1 to 5. I-FABP levels were significantly higher in patients who eventually died from COVID-19 than in survivors, and the optimal I-FABP cut-off point for predicting 28-day mortality was 668.57 pg/mL (sensitivity 0.739, specificity 0.765). Plasma levels of I-FABP, but not citrulline, were associated with significantly higher mortality and appeared to be a predictor of poor outcome in multivariate logistic regression analysis. In conclusion, I-FABP seems to be an effective prognostic marker in critically ill COVID-19 patients. Assessing mortality risk based on intestinal markers may be helpful in making clinical decisions regarding the management of intestinal injury, imaging diagnostics, and potential surgical interventions.

The machine learning modeling process conventionally results in selecting a single model that maximizes a selected performance metric. However, this approach leads to abandoning a more profound analysis of slightly inferior models. Particularly in medical and healthcare studies, where the objective extends beyond predictions to generation of valuable insights, relying solely on a single model can result in misleading or incomplete conclusions. This problem is particularly pertinent when dealing with a set of models known as Rashomon set , with performance close to maximum one. Such a set can be numerous and may contain models describing the data in a different way, which calls for comprehensive analysis. This paper introduces a novel method to explore models in the Rashomon set, extending the conventional modeling approach. We propose the Rashomon_DETECT algorithm to detect models with different behavior. It is based on recent developments in the eXplainable Artificial Intelligence (XAI) field. To quantify differences in variable effects among models, we introduce the Profile Disparity Index (PDI) based on measures from functional data analysis. To illustrate the effectiveness of our approach, we showcase its application on real-world medical problem: predicting survival among hemophagocytic lymphohistiocytosis (HLH) patients – a foundational case study. Additionally, we benchmark our approach on other medical data sets, demonstrating its versatility and utility in various contexts. If differently behaving models are detected in the Rashomon set, their combined analysis leads to more trustworthy conclusions, which is of vital importance for high-stakes applications such as medical applications.

Comorbidities impair the prognosis of diffuse large B-cell lymphoma (DLBCL). Type 2 diabetes mellitus (DMT2) increases the risk of other comorbidities, e.g., heart failure (HF). Thus, we hypothesized that pre-existing DMT2 may negatively affect the outcome of DLBCL. To verify this, DLBCL patients treated with (R)-CHOP were enrolled. 469 patients were eligible, with a median age of 57 years; 356 patients had advanced-stage DLBCL. 126 patients had high-intermediate and 83 high-risk international prognostic index (IPI). Seventy-six patients had DMT2, 46 HF; 26 patients suffered from both DMT2 and HF. In the analyzed group DMT2 or HF significantly shortened overall survival (OS) and progression free survival (PFS): the 5-year OS for patients with DMT2 was 64% vs 79% and for those with HF: 49% vs 79%. The 5-year PFS for DMT2 was 50.6% vs 62.5% and for HF 39.4% vs 63.2%. The relapse/progression incidence was comparable between groups; the non-relapse/progression mortality (NRPM) was significantly higher solely in DMT2 patients (5-year NRPM 22.5% vs 8.4%). The risk of death was higher in patients with higher IPI (HR = 1.85) and with DMT2 (HR = 1.87). To conclude, pre-existing DMT2, in addition to a higher IPI and HF, was a negative predictor for OS and PFS.

Fibronectin (FN) plays an essential role in the host’s response to infection. In previous studies, a significant decrease in the FN level was observed in sepsis; however, it has not been clearly elucidated how this parameter affects the patient’s survival. To better understand the relationship between FN and survival, we utilized innovative approaches from the field of explainable machine learning, including local explanations (Break Down, Shapley Additive Values, Ceteris Paribus), to understand the contribution of FN to predicting individual patient survival. The methodology provides new opportunities to personalize informative predictions for patients. The results showed that the most important indicators for predicting survival in sepsis were INR, FN, age, and the APACHE II score. ROC curve analysis showed that the model’s successful classification rate was 0.92, its sensitivity was 0.92, its positive predictive value was 0.76, and its accuracy was 0.79. To illustrate these possibilities, we have developed and shared a web-based risk calculator for exploring individual patient risk. The web application can be continuously updated with new data in order to further improve the model.

The soluble urokinase-type plasminogen activator receptor (suPAR) is involved in the pathogenesis of acute kidney injury (AKI). Our goal was to establish the optimal suPAR cut-off point for predicting the need for kidney replacement therapy (KRT) use in sepsis patients and to analyze survival rates based on the suPAR level, AKI diagnosis, and the requirement for KRT. In total, 51 septic patients were included (82% septic shock; 96% mechanically ventilated, 35% KRT). Patients were stratified according to the AKI diagnosis and the need for KRT into three groups: AKI(+)/KRT(+), AKI(+)/KRT(−), and AKI(−)/KRT(−). A control group (N = 20) without sepsis and kidney failure was included. Sepsis patients had higher levels of the suPAR than control (13.01 vs. 4.05 ng/mL, p < 0.001). On ICU admission, the suPAR level was significantly higher in the AKI(+)/KRT(+) group than in the AKI(+)/KRT(−) and AKI(−)/KRT(−) groups (18.5 vs. 10.6 and 9.5 ng/mL, respectively; p = 0.001). The optimal suPAR cut-off point for predicting the need for KRT was established at 10.422 ng/mL (area under the curve 0.801, sensitivity 0.889, specificity 0.636). Moreover, patients AKI(+)/KRT(+) had the lowest probability of survival compared to patients AKI(+)/KRT(−) and AKI(−)/KRT(−) (p = 0.0003). The results indicate that the suPAR measurements may constitute an important element in the diagnosis of a patient with sepsis.