Stratifying and predicting progression to acute liver failure during the early phase of acute liver injury

Abstract

Acute liver failure (ALF) is a serious disease that progresses from acute liver injury (ALI) and that often leads to multiorgan failure and ultimately death. Currently, effective treatment strategies for ALF, aside from transplantation, remain elusive, partly because ALI is highly heterogeneous. Furthermore, clinicians lack a quantitative indicator that they can use to predict which patients hospitalized with ALI will progress to ALF and the need for liver transplantation. In our study, we retrospectively analyzed data from 319 patients admitted to the hospital with ALI. By applying a machine-learning approach and by using the SHapley Additive exPlanations (SHAP) algorithm to analyze time-course blood test data, we identified prothrombin time activity percentage (PT%) as a biomarker reflecting individual ALI status. Unlike previous studies predicting the need for liver transplantation in patients with ALF, our study focused on PT% dynamics. Use of this variable allowed us to stratify the patients with highly heterogeneous ALI into six groups with distinct clinical courses and prognoses, i.e. self-limited, intensive care-responsive, or intensive care-refractory patterns. Notably, these groups were well predicted by clinical data collected at the time of admission. Additionally, utilizing mathematical modeling and machine learning, we assessed the predictability of individual PT% dynamics during the early phase of ALI. Our findings may allow for optimizing medical resource allocation and early introduction of tailored individualized treatment, which may result in improving ALF prognosis.

Fig. 1.

Exploring a biomarker for individual ALF progression: A) The ROC curve of RF classifiers trained to predict the need for transplantation based on the blood test data at 7 days postadmission is presented (using data from Fig. S1A). The corresponding ROC–AUC is calculated and displayed at the top of the panel. B) Feature importance of the predictive model in A is illustrated as a SHAP summary plot (using data from Fig. S1A). The y-axis represents the blood test items, arranged in order of their contribution to the prediction. The contribution of each feature for each patient (each point) to the prediction is represented as SHAP values (x-axis). A higher SHAP value means a higher contribution to the likelihood of the need for transplantation, while a lower value indicates a higher contribution to the likelihood of no need. The color of each point in each feature represents the value of that feature for the patient, with higher values shown in red and lower values in blue. PT%, prothrombin time activity percentage; Plt, platelet; Cre, creatinine; BUN, blood urea nitrogen; ALT, alanine aminotransferase; LDH, lactate dehydrogenase; Alb, albumin; D/T-bil, ratio of direct bilirubin (D-bil) to total bilirubin (T-bil); AST, aspartate aminotransferase. C) The time course of observed PT% values for all cases is depicted. The green and orange plots (i.e. left and right panels) represent the TFS (n = 264) and non-TFS (n = 55) patients, respectively (using data from Fig. S1B). D) A logistic regression model for predicting severe patients using the PT% values on day 7 is computed (using data from Fig. S1A). The green and orange dots (i.e., the dots with y-axis values of 0 and 1, respectively) represent the TFS and non-TFS patients, respectively. The red dashed line indicates the PT% threshold value (i.e. PT% = 51.30%).

Fig. 2.

Stratifying and characterizing PT% dynamics during the progression of ALF: A) The time-course patterns of PT% for each group stratified by unsupervised time-series clustering are depicted and colored accordingly (using data from Fig. S1B). The green and orange plots represent the TFS (n = 264) and non-TFS (n = 55) patients. B) The PT% at admission and 7 days postadmission are compared across the stratified groups (using data from Fig. S1B) with colors corresponding to those used in A. The red dashed line indicates the PT% threshold value obtained in Fig. 1D (i.e. PT% = 51.30%). C) Shown is the ROC curve of the binary classification prediction (e.g. either G6 or not) using RF based on the clinical dataset on admission (using data from Fig. S1C) except for G1 and G2. The corresponding ROC–AUCs are calculated and displayed at the top of each panel. D) Feature importance of the predictive model in C is illustrated as a SHAP summary plot (using data from Fig. S1C). Here, only the top 5 features are shown, but the SHAP values for all features are presented in Fig. S4. ATIII, antithrombin III; Che, cholinesterase; PT%, prothrombin time activity percentage; AST, aspartate aminotransferase; gGTP, γ-glutamyl transpeptidase; ALT, alanine aminotransferase; LDH, lactate dehydrogenase; APTT, activated partial thromboplastin time; NH3, serum ammonia; no LA, no liver atrophy.

Fig. 3.

Predicting individual PT% dynamics during the early progression phases: A) The reconstructed individual PT% dynamics for 24 representative patients are displayed. The dots and black solid curves represent the observed PT% data and the best-fitted model by NLMEM based on the entire PT% dataset (i.e. model fitting), respectively (using data from Fig. S1B). B) The predicted PT% dynamics by the mathematical model with RF-predicted parameters based on the blood test data on admission for the 24 patients (i.e. RF prediction) are presented. The green and orange solid lines correspond to TFS and non-TFS cases, respectively, while the shaded area in each panel indicates the mean and 95% prediction interval of the model prediction, respectively (using data from Fig. S1C). The dots and black solid curves are the same as in A. C) The average and individual RMSEs between the observed PT% data and the RF prediction with different datasets are depicted in the left and right panels, respectively (using data from Fig. S1E). Note that $D_0$ represents clinical data on admission, including blood test data; $\mathrm{DT}{\mathrm{I}}_0$ includes $D_0$ and treatment information on admission; $D_t$ includes $D_0$ and blood test data until day t postadmission; $\mathrm{DT}{\mathrm{I}}_t$ includes $D_t$ and treatment information until day t postadmission. The green and orange plots represent TFS and non-TFS patients, respectively.