Lasso Regression-based Model for Cross-Sectional Identification of Significant Hepatic Fibrosis in NAFLD LASSO-based identification of hepatic fibrosis in NAFLD

Non-alcoholic fatty liver disease (NAFLD) is a chronic liver disorder characterized by excessive hepatic fat accumulation and closely associated with metabolic dysfunction. It has become one of the most prevalent liver diseases worldwide, affecting an estimated 38% of adults, a figure that continues to increase alongside global obesity and diabetes epidemics^[1-4]. By 2040, the global prevalence of NAFLD is projected to reach 55.4%^[5]. Consequently, NAFLD has emerged as a leading chronic liver disease and a major contributor to liver-related morbidity and mortality^[6].

The pathological spectrum of NAFLD ranges from simple steatosis to non-alcoholic steatohepatitis (NASH), liver fibrosis, cirrhosis, and hepatocellular carcinoma (HCC)^[7]. The presence and severity of fibrosis are strongly associated with liver-related mortality, cardiovascular events, and overall survival, with prognosis worsening as fibrosis progresses^[8,9]. Although liver biopsy remains the current "gold standard" for the assessment of fibrosis, its invasiveness, procedural risks, and limited patient acceptance restrict its widespread clinical use. Therefore, there is an urgent need for safe, convenient, cost-effective, and accurate noninvasive tools to evaluate liver injury, identify patients with NAFLD with significant fibrosis (SigFib), and guide clinical management.

Several noninvasive scoring systems, including the aspartate aminotransferase-to-platelet ratio index (APRI), fibrosis-4 index (FIB-4), and liver stiffness measurement (LSM), have been used to assess the likelihood of liver fibrosis in NAFLD. However, most of these tests were originally developed for viral hepatitis and have shown limited diagnostic accuracy in NAFLD populations. Therefore, the development of NAFLD-specific models is necessary. This study aimed to establish a practical clinical tool to identify patients with significant fibrosis, facilitate severity stratification, guide individualized management, and construct a noninvasive diagnostic model for SigFib in NAFLD.

This study was conducted at Beijing Ditan Hospital, Capital Medical University. Patients diagnosed with biopsy-confirmed NAFLD between October 2008 and December 2018 were included.

Patients with biopsy-confirmed NAFLD were enrolled for model development and internal validation. The inclusion criteria were as follows: (1) age 18–65 years; (2) complete clinical data, and (3) definite NAFLD diagnosis based on liver tissue. The exclusion criteria were as follows: (1) liver diseases, such as alcoholic hepatitis, autoimmune hepatitis, and drug-induced hepatitis; (2) liver diseases caused by viral infections, including various hepatitis viruses, Epstein-Barr virus (EBV), cytomegalovirus (CMV), and human immunodeficiency virus (HIV); (3) mental disorders; and (4) liver tumors.

All data were obtained from the Hospital Information System (HIS) and Laboratory Information System (LIS). Demographic data, including age, sex, height, weight, body mass index (BMI), and economic status were collected. Data on medical history, lifestyle (smoking, alcohol consumption, diabetes, hypertension, hyperlipidemia, and sedentary habits), and family history (diabetes, hypertension, hyperlipidemia, and hepatocellular carcinoma) were collected. Laboratory tests were conducted to assess liver function, including alanine aminotransferase (ALT), aspartate aminotransferase (AST), and total bilirubin (TBIL), albumin (ALB), alkaline phosphatase (ALP), and cholinesterase (CHE); renal function tests such as uric acid (UA); lipid profile, including total cholesterol (TC), triglycerides (TG), low-density lipoprotein cholesterol (LDL), high-density lipoprotein (HDL), apolipoprotein A1 (apoA1), apolipoprotein B (apoB), and homocysteine (HCY); blood glucose (GLU); blood routine tests such as white blood cell count (WBC), neutrophil count (NEUT), hemoglobin (HGB), and platelet count (PLT); coagulation function tests such as erythrocyte sedimentation rate (ESR) and international normalized ratio (INR); and liver fibrosis serum markers such as procollagen III (PCIII), procollagen type IV (PIVC), hyaluronic acid (HA), and laminin (LN) were collected. Imaging data such as the controlled attenuation parameter (CAP) and LSM were also obtained. Finally, histopathological data from the liver biopsies were assessed, including the degree of steatosis (S0–S3), inflammation grade (G0–G3), and fibrosis stage (F0–F4).

Under ultrasound guidance, liver tissue was obtained using a 16G biopsy needle, with a specimen length ranging from 1.5 to 2.5 cm (minimum 1.0 cm). The specimens were sectioned and stained using routine hematoxylin-eosin, reticulin, and/or Masson’s trichrome. Hepatocyte steatosis (S0–3), inflammation grade (G0–3), and fibrosis stage (F0–4) were evaluated according to the "Guidelines for the Diagnosis and Treatment of Nonalcoholic Fatty Liver Disease" (2006 edition). The patients were classified into non-significant fibrosis (non-SigFib, S0–1) and significant fibrosis (SigFib, S2–4) groups^[10].

Statistical analyses were performed using SPSS 26.0 and R 4.4.3. Categorical variables were compared using Pearson's chi-square or Fisher's exact tests. The normality of continuous variables was assessed using the Kolmogorov-Smirnov test. Non-normally distributed variables (P < 0.05) were summarized as median (interquartile range) and compared using the Mann–Whitney U test, whereas normally distributed variables (P ≥ 0.05) were presented as mean ± standard deviation. The homogeneity of variance was evaluated using Levene's test; if variances were unequal, Welch's t-test was applied.

Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to identify variables associated with SigFib in NAFLD, and a multivariate logistic regression model was subsequently constructed. A nomogram was developed based on the multivariate model. Receiver operating characteristic (ROC) curves were generated using the pROC package, and areas under the ROC curves (AUC) were compared using the DeLong test. Decision curve analysis (DCA) was performed using the rmda package and data visualization was conducted using ggplot2. All tests were two-sided, with a significance level of α = 0.05.

This study was approved by the Ethics Committee of Beijing Ditan Hospital (Ethical ID: Jing Lun Di Zi 2018-052-01). All data were used solely for academic purposes.

A total of 466 NAFLD patients were included, comprising 281 males (60.3%) and 185 females (39.7%). Among them, 354 patients were classified as non-SigFib and 112 as SigFib. Comparative analysis indicated that, relative to the non-SigFib group (F < 2), the SigFib group (F ≥ 2) exhibited the following distinguishing characteristics:

Demographic characteristics: The SigFib NAFLD group had a significantly higher proportion of female patients [54.46% (61/112) vs. 35.03% (124/354), χ² = 13.426, P < 0.001], a higher rate of positive family history of diabetes [35.71% (40/112) vs. 18.08% (64/354), χ² = 15.262, P < 0.001], and a higher prevalence of diabetes [16.07% (18/112) vs. 8.47% (30/354), χ² = 5.314, P = 0.021]; older age [median 46.50 (32.50, 54.00) vs. 38.00 (29.00, 48.25) years, Z = 4.140, P < 0.001] and higher BMI [29.72 (28.54, 30.67) vs. 27.76 (24.86, 29.40), Z = 7.520, P < 0.001]; and worse economic status [72.32% (81/112) vs. 51.41% (182/354), χ² = 15.129, P < 0.001]. No significant differences were observed in the rate of positive family history of hyperlipidemia or the prevalence of hyperlipidemia [47.32% (53/112) vs. 53.11% (188/354), χ² = 1.141, P = 0.286; 32.14% (36/112) vs. 40.11% (142/354), χ² = 2.289, P = 0.130].

Clinical laboratory indicators: The SigFib NAFLD group exhibited higher levels of ESR [median 8.00 (6.00, 15.00) vs. 7.00 (5.00, 12.00), Z = 2.453, P = 0.014], AST [58.50 (34.25, 94.75) vs. 42.00 (28.00, 64.25), Z = 3.721, P < 0.001], and GLU [6.27 (5.47, 7.97) vs. 5.74 (5.29, 6.32), Z = 4.226, P < 0.001]; as well as higher levels of INR [1.03 (0.97, 1.09) vs. 0.99 (0.93, 1.03), Z = 5.147, P < 0.001], PcIII [32.73 (14.59, 52.26) vs. 16.49 (12.08, 25.07), Z = 5.775, P < 0.001], and PIVC [76.46 (65.26, 88.65) vs. 67.34 (56.11, 78.36), Z = 4.011, P < 0.001]. Conversely, the SigFib NAFLD group had significantly lower levels of CHE [9,338.50 (7,863.00, 10,442.25) vs. 9,783.00 (8,592.75, 11,050.00), Z = 2.625, P = 0.009], peripheral neutrophil count [3.08 (2.44, 4.13) vs. 3.48 (2.91, 4.30), Z = 2.837, P = 0.005], HGB [144.00 (135.00, 158.00) vs. 152.00 (139.00, 162.00), Z = 2.472, P = 0.013], and PLT [184.50 (143.50, 234.75) vs. 225.00 (181.75, 262.00), Z = 4.530, P < 0.001].

Imaging indicators: The SigFib NAFLD group exhibited markedly higher LSM values [median 11.35 (8.33, 15.90) vs. 6.60 (5.60, 8.30), Z = 9.968, P < 0.001] and higher CAP values [275.00 (246.75, 310.00) vs. 262.50 (211.75, 289.00), Z = 3.660, P < 0.001]. The patient demographics and clinical characteristics are presented in Table 1.

Variable selection was performed using LASSO regression, with the optimal regularization parameter determined by 10-fold cross-validation. Based on the "one standard error" criterion (lambda.1se = 0.027), 10 variables with non-zero coefficients and significant discriminative value (sex, age, BMI, TBIL, HDL, PLT, INR, PcIII, LSM, GLU) were selected from an initial pool of 39 candidate variables. The LASSO variable selection results are shown in Figure 1.

Figure 1.  LASSO regression analysis diagram for variable selection in the model.

Using stepwise regression, ten independent variables (sex, age, BMI, TBIL, HDL, PLT, INR, PcIII, LSM, GLU) were identified and incorporated into a multivariate logistic regression model. The final model equation is: P(SigFib=1) = 1 / (1 + exp(-(−2.084 + 0.442Sex + 0.485Age + 1.327BMI - 0.668TBIL - 0.629HDL - 0.502PLT + 0.512INR + 0.857PcIII + 0.921LSM + 0.254GLU))).

Female sex [adjusted odds ratio (aOR) = 1.56, 95% CI: 1.07–2.29; Z = 2.29, P = 0.022], older age [aOR = 1.62, 95% CI: 1.10–2.39; Z = 2.42, P = 0.016], higher BMI [aOR = 3.77, 95% CI: 2.48–5.73; Z = 6.22, P < 0.001], higher INR [aOR = 1.67, 95% CI: 1.24–2.24; Z = 3.39, P < 0.001], higher PcIII [aOR = 2.36, 95% CI: 1.73–3.22; Z = 5.46, P < 0.001], higher LSM [aOR = 2.51, 95% CI: 1.74–3.62; Z = 4.87, P < 0.001], and higher GLU levels [aOR = 1.29, 95% CI: 0.96–1.74; Z = 1.69, P = 0.092] were associated with the presence of significant fibrosis in NAFLD. Conversely, higher levels of TBIL [aOR = 0.51, 95% CI: 0.30–0.87; Z = −2.48, P = 0.013], HDL [aOR = 0.53, 95% CI: 0.37–0.76; Z = −3.41, P < 0.001], and PLT [aOR = 0.61, 95% CI: 0.43–0.87; Z = −2.80, P = 0.005] were associated with a lower probability of existing significant fibrosis in NAFLD. The regression coefficients and significance values are presented in Table 2.

A nomogram was developed based on these ten variables to allow individualized assessment of the probability of significant fibrosis in patients with NAFLD (Figure 2).

Figure 2.  Nomogram for Detecting Significant Fibrosis in NAFLD.

The model was internally validated using bootstrap resampling (1,000 repetitions, random seed = 123). The original ROC-AUC was 0.919 (95% CI: 0.894–0.945), with an optimal cutoff value of 0.25, sensitivity of 83.9%, and specificity of 85.3% (Figure 3A). The corrected AUC was 0.907 (optimism = 0.024) and the calibration slope was 1.000 (95% CI: 0.813–1.215). Bootstrap confidence intervals closely followed the reference line, and the Hosmer-Lemeshow test yielded P = 0.845. These results indicate that the model demonstrated excellent discrimination and calibration with no significant deviations (Figure 3B).

Figure 3.  Discrimination and Calibration of Diagnostic Model.

The diagnostic performance of the SigFib model was compared with that of the APRI, LSM, and FIB-4 models for identifying significant liver fibrosis. The AUCs were 0.919 for the SigFib model, 0.681, 0.812, and 0.712 for FIB-4, respectively. DeLong’s test confirmed that the SigFib model outperformed all traditional indicators (all P < 0.001), demonstrating its superior discriminative ability and potential clinical utility for noninvasive diagnosis of significant fibrosis in NAFLD (Figure 4).

Figure 4.  Comparison of different models.

The AUC values for the SigFib model were 0.919, 0.681, 0.812, and 0.712 for FIB-4, respectively. Statistically significant differences were observed between the SigFib model and APRI, LSM, and FIB-4 index (P < 0.001).

Across a wide range of threshold probabilities, the SigFib model consistently provided higher net clinical benefit than the reference strategies of “treat all” and “treat none.” These results indicate that the clinical decisions guided by the SigFib model offer superior net benefits across diverse scenarios, highlighting its substantial clinical utility. The DCA for the SigFib model is shown in Figure 5.

Figure 5.  Decision curve of the SigFib NAFLD diagnostic model.

In this study, we developed a multivariate logistic regression model to identify significant fibrosis in NAFLD based on ten key variables using LASSO regression. The model demonstrated excellent discrimination (AUC = 0.919; bootstrap-corrected AUC = 0.907) and good calibration (Hosmer-Lemeshow test, P = 0.845). It significantly outperformed the APRI, FIB-4, and LSM alone. By integrating LSM with routine serological markers, the model can be applied to centers equipped with FibroScan. LASSO regression was employed to minimize overfitting and enhance model stability.

Our findings highlight the pivotal role of metabolic factors in the pathogenesis of NAFLD-related fibrosis. Specifically, female sex, older age, and higher BMI were identified as independent factors associated with significant fibrosis, consistent with previous studies highlighting the strong associations between obesity, insulin resistance, sex differences, and NAFLD progression^[11-15].

The discriminative performance of the model was enhanced by incorporating key hematological, biochemical, and imaging markers. Reduced PLT levels and elevated INR reflect impaired liver functional reserves and progression of fibrosis^[16,17]. Elevated PcIII levels indicate active hepatic collagen deposition and correlate significantly with the histological stage of fibrosis, serving as a marker of fibrogenesis^[18,19]. Lower HDL levels may reflect dysregulated hepatic lipid metabolism and chronic inflammation^[20-22]. The LSM, an imaging marker of liver stiffness, was identified as a significant factor. Integrating these variables allowed the model to outperform individual indicators in distinguishing between significant and nonsignificant fibrosis.

Compared with traditional scoring systems, the model developed in this study demonstrated clear advantages. APRI and FIB-4 were originally derived from populations with viral hepatitis and showed limited diagnostic accuracy for NAFLD^[23]. Although LSM is widely used for noninvasive assessments, its performance can be affected by obesity or concomitant inflammation^[24]. By integrating clinical characteristics, laboratory markers, and imaging parameters, the present model overcomes the limitations of individual indicators and improves discriminative ability.

DCA was performed to evaluate the clinical utility of this model. The results demonstrated that the model provides substantial net benefit across a wide range of decision thresholds, indicating that it is not only statistically robust but may also offer practical value in clinical decision-making, guiding stratified management, and intervention according to disease severity.

Previous studies largely relied on serological scores (APRI and FIB-4) or imaging-based elastography measures such as LSM alone. In contrast, the present study systematically integrated LSM with routine serological markers to identify significant hepatic fibrosis in patients with NAFLD. This model may serve as a noninvasive screening tool, potentially reducing unnecessary liver biopsies in low-risk patients.

However, this study had several limitations. First, this was a single-center retrospective study with a relatively small sample size, which may have introduced selection bias. Second, the model was validated internally, and its generalizability requires confirmation in independent external cohorts. Third, some emerging noninvasive markers, such as metabolism-related biomarkers or radiomic features, were not included, which may further enhance the discriminative performance in future studies. Importantly, this model is not intended to replace FibroScan or expand its accessibility; rather, by integrating routinely available serological markers (e.g., those included in the APRI and FIB-4) with existing LSM values, it improves the ability to discriminate between significant and nonsignificant hepatic fibrosis.

In conclusion, the SigFib diagnostic model developed in this study using LASSO regression demonstrated robust performance in distinguishing significant fibrosis in NAFLD, surpassing traditional scoring systems and showing substantial clinical utility. Future studies should focus on external validation in multicenter, large-scale prospective cohorts and investigate the incorporation of emerging technologies, such as metabolomics, proteomics, and AI-based imaging analysis, to further improve the accuracy and generalizability of NAFLD fibrosis assessments.