In this study (NCT04295330), we performed a secondary analysis on the prospective clinical trials (NCT05492669) in Medical Center: West China Hospital (Sichuan, China) (14) and screened 130 patients suffering from primary liver cancer who underwent elective hepatectomy from 2019 to 2021. Study eligibility included those undergoing elective hepatectomy for primary liver cancer, aged 18–80 years, American Society of Anesthesiologists (ASA) physical status I–III; with follow-up assessment for NRS. All clinical data, including grading of NRS and clinical and physiological parameters were prospectively collected. We excluded patients without follow-up and those body weight <40 or >100 kg; metastases occurring in other distant organs; severe hepatic insufficiency (aspartate aminotransferase or alanine transaminase or bilirubin >2.5 times the upper limit of normal), renal impairment (creatinine clearance <60 ml/min); cardiac rhythm disorders or systolic heart failure (second-and third-degree heart block, ejection fraction <50%); with allergies to any of the trial drugs; chronic opioid use; inability to comprehend numeric rating scale.

Anesthesia procedures were performed in accordance with standardized criteria. For premedication, patients received 50 mg flurbiprofen axetil by intravenous infusion 30 min before the surgery initiation. Intravenous access to the upper limb was performed and standard monitoring was performed, including electrocardiography, pulse oximetry, blood pressure, and the bispectral index (BIS). Then anesthesia was induced by propofol 1.5–2 mg/kg, midazolam 2 mg and sufentanil 0.2–0.3 μg/kg. Tracheal intubation was facilitated with cis-atracurium 0.2 mg/kg. Whole anesthesia progress was maintained with remifentanil 0.1–0.2 μg/kg/min and desflurane or sevoflurane in a mixture of air 40% and O₂ 60% to maintain BIS within 40–60. Systolic arterial blood pressure was maintained within 20% of baseline values; when hypotension (MAP <55 mmHg) occurred, patients were treated with intravenous phenylephrine. The incision was infiltrated with 0.25% ropivacaine 20 ml at the end of the procedure. Following the completion of the surgical procedure, the patients were extubated in the operating room and admitted to the post-anesthesia care unit (PACU).

Observations of pain were conducted every 10 min using the numerical rating scale (NRS), and those who scored higher were treated with sufentanil 0.1 µg/kg every 10 min. Aldrete scores were recorded prior to leaving the PACU. It was only possible to discharge patients from the PACU if the Aldrete score was at least 9 and there was no evidence of pain or PONV (15).

Patients received intravenous analgesia with PCIA for the first 72 h post-operatively. In case pain occurs, patient may press the PCIA button repeatedly until feeling relief. Mephedrone 5 mg would be administered if the NRS score of the participant was >3 after taking the maximum dose (10 ml/h).

The day before the operation, all patients received an explanation of how to rate pain intensity on NRS, identifying 0 as “no pain” and 10 as “worst imaginable pain”. In the final assessment of NRS grade, face-to-face follow-ups were conducted 1 day after surgery and telephone follow-ups were conducted 3 months after surgery. Acute postoperative pain is defined as the pain that occurs during movement (e.g., deep breathing) at postoperative 24 h, and CPSP as the pain that occurs at 3 months after surgery.

In the above 3-months follow-ups, the patients suffering from CPSP also received an explanation of ID pain (16). The items included the following: (1) “Did the pain feel like pins and needles?” (2) “Did the pain feel hot/burning?” (3) “Did the pain feel numb?” (4) “Did the pain feel like electrical shocks?” (5) “Is the pain made worse with the touch of clothing or bed sheets?” and (6) “Is the pain limited to your joints?” “Yes” answers to Questions 1–5 were scored as 1, whereas a “yes” answer to Question 6 was scored as −1. Higher scores suggested a neuropathic component to the pain.

Blood samples are collected at 24 h after surgery for subsequent measurement, which are labelled, centrifuged, frozen, and stored locally at −80° in line with experimental standards. Levels of 8 plasma biomarkers including VEGF, IL-1β, IL-6, CXCL10, TNF-α, MMP2, MMP3 and MMP9 are measured using Human Premixed Multi-Analyte Kit (R&D, Cat: LXSAHM-15, Lot: L141228).

Permuted importance approach in random forest was applied for elimination of some variables that were not predictive for CPSP to improve model accuracy. Permuted importance for a particular variable was calculated by comparing the difference in the prediction accuracy of the out-of-bag data to the prediction accuracy when the variable was noised up by randomly permuting its values. The higher the predictive power variable obtained, the larger importance value it earned. Variables were ranked from high to low according to their permuted importance score and those with the highest scores were selected. Variable selection of this study was performed using R, version 3.6.1.

Through the building and cross-validating of predictive models to classify CPSP risk, multiple machine learning algorithms to identify important biomarkers and clinical variables can be employed. After comparison, Support Vector Machines (SVM) and Random Forest were selected. Model building were performed using R, version 3.6.1.

Participants were selected from a clinical trial recruiting between June 2019 and June 2021. Participant flow through the study was shown in Figure 1. After the initial screening, 130 patients were deemed eligible to participate in the study. Participants with missing data were removed. Of these, 91 (70.0%) patients completed the blood sample collected and follow-up at 3 months, while radiological examinations were performed to exclude recurrence of the primary tumor, distant metastases from the primary tumor, and new tumors. Thus, this report presents results of 91 participants who were included in the final analysis.

91 individuals undergoing hepatectomy were included in the analyses. 70 (77.0%) were male; 21 (23.0%) were female. The mean ± SD. age of all participants was 54.1 ± 11.0 years (range 27–78 years). The average duration of surgery was 3.2 ± 1.2 h, and the average duration of anesthesia was 4.4 ± 1.2 h. The average volume of bleeding was 289.8 ml, which varying from 10 to 1,200 ml. Open hepatectomy was performed in 76 patients (83.51%), while laparoscopic hepatectomy was performed in 15 patients (16.49%). Of these, the incision length varied from 23 to 30 cm for open procedures, with an average of 26.82 cm.

All participants reported having acute postoperative pain. Of these, 33 (36.2%) reported mild pain (NRS 1–3), 46 (50.6%) reported moderate pain (NRS 4–6), and 12 (13.2%) reported severe pain (NRS > 6).

There were 49 (53.8%) participants who reported no pain 3 months after surgery, 40 (44.0%) participants who reported mild pain, and 2 (2.2%) participants who reported moderate pain. Pain distribution and characteristics were shown in Figures 2A,B.

More recently, the six-item ID Pain was developed as a brief, self-administered screening tool for detecting neuropathic pain in primary care settings (16). Among those who reported CPSP, 12 scored 0 (28.6%), indicating they were less likely to be neuropathic; while 22 scored 1 (52.3%) and 8 scored 2 (19.1%), indicating that most patients suffer from CPSP associated with neuropathic pain. The results of the ID pain distribution were shown in Figure 2C.

To explore the effect of each biomarker and clinical variables, random forest was performed for them shown in Figure 3. NRS (13.14) was significantly associated with the risk of CPSP. Also, CXCL10 (4.24), MMP3 (2.53), TNF-α (2.10), IL-1β (2.27), MMP2 (2.14), Anesthesia Duration (2.20), Operation Duration (2.22) and VEGF (4.96) were also significantly related to the risk of CPSP (threshold = 2).

91 patients were randomized into two independent sets: 64 patients for model building and cross validation in train set (Set 1), and 27 patients in test set (Set 2). Two groups share similar physical and clinical characteristics.

SVM model for CPSP risk were first trained and tested in Set1 by cross validation. The model with best performance was selected, and this model included 2 features from biomarkers and clinical variables of our interest. NRS at the first postoperative day seemed to be the most significant one as identified by random forest shown in Figure 3. In line with this, previous studies attached great importance to postoperative acute pain towards the onset and progress of CPSP (17, 18). As mentioned above, we finally picked it as a representative clinical variable. CXCL10, MMP3, TNF-α, IL-1β, MMP2 and VEGF were also taken into account in this model not only for its remarkable importance, but for indispensable role in neuro-inflammatory pathways. For better visual comparison, models including CXCL10, MMP3, VEGF, TNF-α, IL-1β, MMP2 and NRS were shown in Figure 4 to compare their performances. The final SVM model with radial kernel included NRS and MMP3 as model features, while cost = 2, gamma = 0.01 as model parameters. The classifying performance measures in Set 1 included: accuracy = 0.890, sensitivity = 0.963, specificity = 0.889, AUC = 0.953. Figure 5 showed the predicted results of this final SVM model for Set 1 patients, 32/36 patients actually had CPSP when they were predicted to have CPSP, while 25/27 patients without CPSP were predicted to not develop CPSP. Matrix metalloproteinases (MMPs) have been implicated in the modulation of synaptic plasticity, glial activation, and long-term potentiation in the CNS and MMP3 was proved to regulate of nociceptive processing in thermal hyperalgesia and tactile hypersensitivity (19), which was consistent with the ID pain result above that most CPSP associated with neuropathic pain.

The final SVM model was validated in Set 2. It predicted CPSP risk with accuracy = 0.857, sensitivity = 0.800, specificity = 0.923, AUC = 0.892. The validated results are shown in Figure 6: 12/15 patients actually had CPSP when they were predicted to have CPSP, while 12/13 patients without CPSP were predicted to not develop CPSP.

The performances of the final SVM classifier were compared with that of random forest model (also built in Set 1 and tested in Set 2). As shown in Figure 7, the ROC curve (AUC = 0.917) of the final SVM classifier was not only higher than that of the SVM classifiers with other features of our interest as described above Set 1, but also higher than random forest model (AUC = 0.783).