INTRODUCTION
Healthcare expenditures among patients with chronic spinal conditions and those undergoing surgery for management continue to grow at a notable rate.[ 1 ] With a significant burden on quality of life, surgical intervention remains a critical and often unavoidable pillar of treatment paradigms for these individuals. A considerable portion of these expenses stems from postoperative complications, continued rehabilitation and care, and medication; Kammien et al . highlighted that emergency department visits, in addition to the traditionally accepted metric of hospital readmission, are important to consider among these patients.[ 2 , 3 , 4 ] However, such factors cannot be viewed in isolation; medical history, surgery type, and characteristics of preoperative management have all been cited as individual predictors of longitudinal resource utilization.
Identifying patients at risk for increased postoperative healthcare utilization is therefore critical for optimizing care outcomes and resource allocation. As the number of individuals undergoing cervical spine surgery grows to well over 150,000 a year, the demand for systems capable of accurate prognostication also increases.[ 5 , 6 ] Although traditional statistical methods have been used to weigh the prognostic significance of patient and care plan features, they struggle to capture their relative importance in shaping recovery trajectories. Machine learning (ML) offers more sophisticated techniques capable of synthesizing and analyzing greater volumes of patient data and providing case-specific forecasts of postoperative outcomes. When compared to traditional clinical risk scores such as the American Society of Anesthesiologists (ASA) classification, ML models permit more accurate and deeper insight into adverse event prediction.[ 7 ] Several groups have highlighted their superiority over classic comorbidity indices in predicting short-term outcomes following cervical spine surgery, such as prolonged length of stay, discharge to rehabilitation, and 30- and 90-day readmission, and overall functional recovery.[ 8 , 9 , 10 , 11 , 12 ]
Although prior studies have individually looked at several patient features and endpoints, few have focused on the prediction of healthcare utilization among a comprehensive cohort of cervical spine surgery patients. These factors not only serve as drivers of elevated healthcare costs but also as potential proxies for surgical care quality and patient satisfaction. True precision medicine requires better characterization of what attributes are most predictive of both optimal and suboptimal care. The aim of our study is to leverage a range of preoperative patient characteristics – including comprehensive medical, surgical, and social histories – to develop ML models capable of granular prediction of postoperative healthcare utilization among cervical spine surgery patients. We further aim to capture and rank the most highly important features for prediction, enabling patient-centered interventions.
MATERIALS AND METHODS
Data processing
We extracted all degenerative cervical fusion procedures performed on adult patients from 2002 to 2022 using our institutional data warehouse. Identification of relevant cases was carried out using a targeted set of CPT and ICD codes [ Supplemental Table 1 ]. For each case, Electronic Medical Record data and associated codes were used to derive key patient-level variables [ Supplemental Table 2 ]. The principal outcome assessed was high healthcare utilization 90-day postoperatively. Considering the absence of a gold-standard definition for high healthcare utilization, we defined this based on a combination of increased specialist involvement (>75 th percentile office visits to a neurosurgeon or orthopedist or >75 th percentile number of specialist referrals) and high resource use (1+ emergency department or urgent care visit, >75 th percentile pain medication prescriptions, >75 th percentile spine imaging orders). These criteria were determined by prior analyses of healthcare utilization patterns conducted by our group. This aggregate outcome was designed to prioritize sensitivity in identifying patients at higher risk.
Supplemental Table 1
Cohort generation criteria
CPT - Current procedural terminology; ICD - International classification of diseases
| Inclusion CPT | Exclusion CPT | Exclusion ICD | |
| Cervical decompression and fusion | ACDF: ‘22551’ OR (‘22554’ AND ‘63075’) PCDF: ‘22600’ OR ‘22595’ OR ‘22590’ | Tumor: “63275”, “63276”, “63277”, “63278”,”63280”, “63281”, “63282”, “63283”, “63285”, “63286”,”63287”, “63290”, “63300”, “63301”, “63302”, “63303”,”63304”, “63305”, “63306”, “63307”, “22112”, “22114”,”22101”, “22102”, “22100”, “22110” Trauma: “22310”, “22315”, “22318”, “22319”, “22325”, “22326”, “22327” Infection: “63172”,”63173”,”22010”,”22015” Deformity “22206”, “22207”, “22208”, “22210”, “22212”, “22214”, “22216”, “22220”, “22222”, “22224”, “22226”, “22800”, “22802”, “22804”, “22808”, “22810”, “22812”, “22818”, “22819” | Tumor: “C412”, “C414”, “C720”, “C721”, “D166”, “D168”, “D321”, “D434”, “D334”, “C7979”, “C701” Trauma: “S220”, “S239”, “S230”, “S231”, “S240”, “S241”, “S242”, “S120”, “S130”, “S131”, “S134”, “S140”, “S141”, “S142”, “S320”, “S321”, “S322”, “S330”, “S331”, “S335”, “S340”, “S341”, “S342”, “S343”, “G82”, “G83”, “M483”, “S151”, “S121”, “S122”, “S123”, “S124”, “S125”, “S126” Infection: “G061”, “G062”, “M462”, “M463”, “M464”, “M465” Deformity: “M41”,”M40”,”M482” |
Supplemental Table 2
Variable list and definitions
BMI - Body mass index; ICD - International Classification of Diseases; PVD - Peripheral vascular disease; ED/UC - Emergency departments/Urgent cares; NSAIDs - Nonsteroidal Anti-Inflammatory Drugs; ACE - Angiotensin-converting enzyme; HIV - Human immunodeficiency virus; AIDS - Acquired immunodeficiency syndrome; COPD - Chronic obstructive pulmonary disease; FBSS - Failed back surgery syndrome
| Category | Variable | Data type | Coding description |
| Demographics | Age at procedure | Numeric | Date of surgery – Birthdate rounded to lower year |
| Gender | Binary | Self-reported gender | |
| BMI | Numeric | Nearest reported weight and height before surgery | |
| Medical history | Obesity | Binary from Medical History ICD Codes Classifications based on Quan et al. | “E66” or BMI>30 |
| AIDS or HIV | “B20”, “B21”, “B22”, “B24” | ||
| Lymphoma | “C81”, “C82”, “C83”, “C84”, “C85”, “C88”, “C96”, “C900”, “C902” | ||
| Paralysis | “G041”, “G114”, “G801”, “G802”, “G81”, “G82”, “G830”, “G831”, “G832”, “G833”, “G834”, “G839” | ||
| Psychosis | “F20”, “F22”, “F23”, “F24”, “F25”, “F28”, “F29”, “F302”, “F312”, “F315” | ||
| Depression | “F204”, “F313”, “F314”, “F315”, “F32”, “F33”, “F341”, “F412”, “F432” | ||
| Drug abuse | “F11”, “F12”, “F13”, “F14”, “F15”, “F16”, “F18”, “F19”, “Z715”, “Z722” | ||
| Weight loss | “E40”, “E41”, “E42”, “E43”, “E44”, “E45”, “E46”, “R634”, “R64” | ||
| Coagulopathy | “D65”, “D66”, “D67”, “D68”, “D691”, “D693”, “D694”, “D695”, “D696” | ||
| Alcohol abuse | “F10”, “E52”, “G621”, “I426”, “K292”, “K700”, “K703”, “K709”, “T51”, “Z502”, “Z714”, “Z721” | ||
| Liver disease | “B18”, “I85”, “I864”, “I982”, “K70”, “K711”, “K713”, “K714”, “K715”, “K717”, “K72”, “K73”, “K74”, “K760”, “K762”, “K763”, “K764”, “K765”, “K766”, “K767”, “K768”, “K769”, “Z944” | ||
| Renal failure | “I120”, “I131”, “N18”, “N19”, “N250”, “Z490”, “Z491”, “Z492”, “Z940”, “Z992” | ||
| Hypothyroidism | “E00”, “E01”, “E02”, “E03”, “E890” | ||
| Valvular cardiac disease | “A520”, “I05”, “I06”, “I07”, “I08”, “I091”, “I098”, “I34”, “I35”, “I36”, “I37”, “I38”, “I39”, “Q230”, “Q231”, “Q232”, “Q233”, “Z952”, “Z953”, “Z954” | ||
| Blood loss anemia | “D500” | ||
| Deficiency anemia | “D508”, “D509”, “D51”, “D52”, “D53” | ||
| Metastatic cancer | “C77”, “C78”, “C79”, “C80” | ||
| Cardiac arrhythmia | “I441”, “I442”, “I443”, “I456”, “I459”, “I47”, “I48”, “I49”, “R000”, “R001”, “R008”, “T821”, “Z450”, “Z950” | ||
| Rheumatoid arthritis | “L940”, “L941”, “L943”, “M05”, “M06”, “M08”, “M120”, “M123”, “M30”, “M310”, “M311”, “M312”, “M313”, “M32”, “M33”, “M34”, “M35”, “M45”, “M461”, “M468”, “M469” | ||
| Complicated diabetes | “E102”, “E103”, “E104”, “E105”, “E106”, “E107”, “E108”, “E112”, “E113”, “E114”, “E115”, “E116”, “E117”, “E118”, “E122”, “E123”, “E124”, “E125”, “E126”, “E127”, “E128”, “E132”, “E133”, “E134”, “E135”, “E136”, “E137”, “E138”, “E142”, “E143”, “E144”, “E145”, “E146”, “E147”, “E148” | ||
| Uncomplicated diabetes | “E100”, “E101”, “E109”, “E110”, “E111”, “E119”, “E120”, “E121”, “E129”, “E130”, “E131”, “E139”, “E140”, “E141”, “E149” | ||
| Congestive heart failure | “I099”, “I110”, “I130”, “I132”, “I255”, “I420”, “I425”, “I426”, “I427”, “I428”, “I429”, “I43”, “I50”, “P290” | ||
| Complicated hypertension | “I11”, “I12”, “I13”, “I15” | ||
| COPD | “I278”, “I279”, “J40”, “J41”, “J42”, “J43”, “J44”, “J45”, “J46”, “J47”, “J60”, “J61”, “J62”, “J63”, “J64”, “J65”, “J66”, “J67”, “J684”, “J701”, “J703” | ||
| Solid tumor without metastases | “C00”, “C01”, “C02”, “C03”, “C04”, “C05”, “C06”, “C07”, “C08”, “C09”, “C10”, “C11”, “C12”, “C13”, “C14”, “C15”, “C16”, “C17”, “C18”, “C19”, “C20”, “C21”, “C22”, “C23”, “C24”, “C25”, “C26”, “C30”, “C31”, “C32”, “C33”, “C34”, “C37”, “C38”, “C39”, “C40”, “C41”, “C43”, “C45”, “C46”, “C47”, “C48”, “C49”, “C50”, “C51”, “C52”, “C53”, “C54”, “C55”, “C56”, “C57”, “C58”, “C60”, “C61”, “C62”, “C63”, “C64”, “C65”, “C66”, “C67”, “C68”, “C69”, “C70”, “C71”, “C72”, “C73”, “C74”, “C75”, “C76”, “C97” | ||
| Uncomplicated hypertension | “I10” | ||
| Other neurological disorder | “G10”, “G11”, “G12”, “G13”, “G20”, “G21”, “G22”, “G254”, “G255”, “G312”, “G318”, “G319”, “G32”, “G35”, “G36”, “G37”, “G40”, “G41”, “G931”, “G934”, “R470”, “R56” | ||
| PVD | “I70”, “I71”, “I731”, “I738”, “I739”, “I771”, “I790”, “I792”, “K551”, “K558”, “K559”, “Z958”, “Z959” | ||
| Pulmonary circulation disorders | “I26”, “I27”, “I280”, “I288”, “I289” | ||
| Fluid and electrolyte disorders | “E222”, “E86”, “E87” | ||
| Peptic ulcer disease | “K257”, “K259”, “K267”, “K269”, “K277”, “K279”, “K287”, “K289” | ||
| Current tobacco use | “ F17200 ”, “ F17203 ”, “ F17208 ”, “ F17209 ”, “ F17210 ”, “ F17213 ”, “ F17218 ”, “ F17219 ”, “ F17220 ”, “ F17223 ”, “ F17228 ”, “ F17229 ”, “ F17290 ”, “ F17293 ”, “ F17298 ”, “ F17299 ”, “ Z720” | ||
| Former tobacco use | “ F17221 ”, “ F17201 ”, “ F17211 ”, “ F17291 ”, “ Z87891 ” | ||
| Surgical history | Cervical fusion | Binary from Internal Surgical History Coding System | “SHX112” |
| Cervical surgery | “SHX587”, “SHX588”, “SHX589”, “SHX94”, “SHX98” | ||
| Thoracolumbar fusion | “SHX111” | ||
| Thoracolumbar surgery | “SHX699”, “SHX700”, “SHX701”, “SHX801”, “SHX802”, “SHX95”, “SHX96”, “SHX99” | ||
| Unspecified spine fusion | “SHX136”, “SHX223” | ||
| Unspecified spine surgery | “SHX113”, “SHX114”, “SHX136”, “SHX140”, “SHX219”, “SHX398”, “SHX97” | ||
| Preop spine diagnoses | Lumbar stenosis | Binary | “M4806” |
| Lumbar spondylolisthesis | “M4316” | ||
| Lumbar disc disorders | “M5106”, “M5126”, “M5136”, “M5186”, “M51A1”, “M51A2”, “M5146”, “M5116” | ||
| Cervical disc disorders | “M50” | ||
| Cervical stenosis | “M4802”, “M4801” | ||
| Cervical spondylolisthesis | “M4312”, “M4311” | ||
| Cervical disc herniation | “M5020” | ||
| Lumbar disc herniation | “M5126” | ||
| Prior FBSS | “M961” | ||
| Prior pseudoarthrosis | “M960” | ||
| Preoperative medication use | ACE inhibitors | Binary Defined by a prescription or order placed for these medications during the 90 day preoperative windoww | “captopril”, “lisinopril”, “enalapril”, “ramipril”, “benazepril”, “fosinopril”, “moexipril”, “perindopril”, “quinapril”, “trandolapril” |
| ARB | “losartan”, “valsartan”, “azilsartan”, “candesartan”, “eprosartan”, “irbesartan”, “olmesartan”, “telmisartan” | ||
| Antidepressants | “citalopram”, “escitalopram”, “paroxetine”, “fluoxetine”, “fluvoxamine”, “sertraline”, “desvenlafaxine”, “duloxetine”, “levomilnacipran”, “milnacipran”, “tofenacin”, “venlafaxine”, “vilazodone”, “vortioxetine”, “etoperidone”, “trazodone”, “nefazodone”, “reboxetine”, “viloxazine”, “amitriptyline”, “butriptyline”, “clomipramine”, “desipramine”, “dosulepin”, “imipramine”, “iprindole”, “lofepramine”, “melitracen”, “nortriptyline”, “protriptyline”, “trimipramine”, “opipramol”, “tianeptine”, “amoxapine”, “maprotiline”, “mianserin”, “mirtazapine”, “setiptiline”, “isocarboxazid”, “phenelzine”, “tranylcypromine”, “selegiline”, “metralindole”, “moclobemide”, “pirlindole”, “toloxatone” | ||
| Beta-2 agonists | “bitolterol”, “fenoterol”, “isoproterenol”, “levalbuterol”, “pirbuterol”, “procarterol”, “albuterol”, “arformoterol”, “bambuterol”, “clenbuterol”, “formoterol”, “salmeterol”, “abediterol”, “carmoterol”, “indacaterol”, “olodaterol”, “vilanterol” | ||
| Beta blockers | “acebutolol”, “atenolol”, “betaxolol”, “bisoprolol”, “carteolol”, “carvedilol”, “esmolol”, “labetalol”, “metoprolol”, “nadolol”, “nebivolol”, “penbutolol”, “pindolol”, “propranolol”, “sotalol”, “timolol” | ||
| Benzodiazepines | “alprazolam”, “chlordiazepoxide”, “clonazepam”, “clorazepate”, “diazepam”, “estazolam”, “flurazepam”, “halazepam”, “lorazepam”, “oxazepam”, “prazepam”, “quazepam”, “temazepam”, “triazolam” | ||
| Immunosuppressants | “mycophenolate”, “etanercept”, “adalimumab”, “azathioprine”, “cyclosporine”, “tacrolimus”, “sirolimus”, “infliximab”, “natalizumab”, “methotrexate”, “certolizumab”, “prednisone”, “dexamethasone”, “cortisone”, “hydrocortisone”, “prednisolone”, “leflunomide”, “cyclophosphamide”, “chlorambucil”, “budesonide”, “everolimus”, “abatacept”, “basiliximab”, “daclizumab”, “muromonab”, “vedolizumab”, “ustekinumab”, “tocilizumab”, “secukinumab”, “rituximab”, “ixekizumab”, “golimumab”, “anakinra” | ||
| NSAIDs | “celecoxib”, “diclofenac”, “diflunisal”, “etodolac”, “flurbiprofen”, “ibuprofen”, “ketoprofen”, “meloxicamnabumetone”, “naproxen”, “piroxicam”, “sulindac”, “ibuprofen”, “indomethacin”, “ketorolac” | ||
| Opioids | “hydromorphone”, “oxycodone”, “hydrocodone”, “fentanyl”, “methadone”, “morphine”, “codeine”, “butorphanol”, “nalbuphine”, “tramadol”, “tapentadol”, “meperidine”, “buprenorphine”, “oxymorphone” | ||
| Antipsychotics | “chlorpromazine”, “thioridazine”, “mesoridazine”, “levomepromazine”, “loxapine”, “molindone”, “perphenazine”, “thiothixene”, “trifluoperazine”, “haloperidol”, “fluphenazine”, “droperidol”, “zuclopenthixol”, “prochlorperazine”, “amisulpride”, “aripiprazole”, “asenapine”, “blonanserine”, “clotiapine”, “clozapine”, “iloperidone”, “lurasidone”, “mosapramine”, “olanzapine”, “paliperidone”, “perospirone”, “quetiapine”, “remoxipride”, “risperidone”, “sertindole”, “sulpiride”, “ziprasidone”, “zotepine” | ||
| Neuromodulators | “gabapentin”, “pregabalin”, “amitriptyline”, “phenytoin”, “sodiumvalproate”, “lamotrigine”, “carbamazepine”, “levetiracetam”, “oxcarbazepine”, “tiagabine”, “topiramate”, “bupropion” | ||
| Bisphosphonates | “risedronate”, “alendronate”, “ibandronate”, “zoledronicacid”, “pamidronate”, “etidronate” | ||
| Loop diuretics | “bumetanide”, “ethacrynicacid”, “furosemide”, “torsemide” | ||
| Thiazide diuretics | “chlorothiazide”, “chlorthalidone”, “hydrochlorothiazide”, “indapamide”, “metolazone” | ||
| Cinacalcet | “cinacalcet” | ||
| Insulin | “insulin glulisine”, “insulin aspart”, “insulin lispro u-100/u-200”, “regular insulin”, “nph insulin”, “insulindetemir”, “insulin u-100”, “insulin glargine”, “insulin degludec u-100/u-200”, “insulin aspart protamine”, “insulin isophane” | ||
| Oral diabetes meds | “acarbose”, “miglitol”, “metformin”, “alogliptin”, “linagliptin”, “saxagliptin”, “sitagliptin”, “nateglinide”, “repaglinide”, “canagliflozin”, “dapagliflozin”, “empagliflozin”, “ertugliflozin”, “bexagliflozin”, “glimepiride”, “glipizide”, “tolazamide”, “tolbutamide”, “glyburide”, “chlorpropamide”, “rosiglitazone”, “pioglitazone”, “semaglutide”, “colesevelam”, “bromocriptine” | ||
| Calcium supplementation | “calcium carbonate”, “calcium citrate”, “calcium gluconate”, “calcium lactate”, “calcium” | ||
| Vitamin D supplementation | “ergocalciferol”, “colecalciferol”, “dihydrotachysterol”, “doxercalciferol”, “calcifediol”, “calcitriol”, “alfacalcidol”, “paricalcitol” | ||
| Operation time (h) | Integer | ||
| Multi versus single level surgery | Binary | “63015”, “63016”, “63017”, “63035”, “63043”, “63044”, “63048”, “63057”, “63066”, “63076”, “63078”, “22585”, “22810”, “22812”, “22534”, “22614”, “22800”, “22802”, “22804”, “22632”, “22634”, “22552”, “22614”, “22843”, “22844”, “22846”, “22847”, “22854”, “22208”, “22216”, “22226”, “63082”, “63086”, “63088”, “63091”, “63103”, “22858”, “63050”, “63051”, “22512”, “22515”, “22552”, “22846”, “22847”, “63053” | |
| Anterior approach | Binary | “22551”OR (“22554” AND “63075”) | |
| Posterior Approach, C2 to Occiput | Binary | “22590”, “22595” | |
| Posterior Approach, Below C2 | Binary | “22600” | |
| Instrumentation | Binary | “22586”, “27279”, “27280”, “22840”, “22841”, “22842”, “22843”, “22844”, “22845”, “22846”, “22847”, “22849”, “22853”, “22854”, “22859”, “63051”, “22513”, “22514”, “20650”, “22845”, “22846”, “22847” | |
| General anesthesia | Binary | ||
| Inpatient | Binary (0=outpatient) | ||
| Outcome | 90-day High resource utilization | Binary | Aggregate variable of spine imaging, pain medication orders, ED/UC visits, spine surgeon office visits, and referrals |
Model development
We built separate predictive models for cervical fusion and decompression using stratified five-fold cross-validation, repeated across ten independent iterations. Four Python classification models were used: Elastic Net, Support Vector Machine, Balanced Random Forest (BRF), and Extreme Gradient Boosted Tree. In addition, a six-layer neural network was implemented using TensorFlow. To address the issue of class imbalance, inverse class weighting was used when appropriate. For models supporting early stopping, 10% of the dataset was randomly reserved. Missing values were handled using Multivariate Imputation by Chained Equations by using the MiceForest package in Python. Model hyperparameters were fine-tuned through a nested five-fold cross-validation framework using the Ray TuneSearchCV algorithm [ Supplemental Table 3 ].
Supplemental Table 3
Hyperparameter search spaces
RF - Random forest; ENet - Elastic net; XGB - Extreme gradient boosted; SVC - Support vector classification
| Model | Parameter | Description | Search space |
| RF | n_estimators | Number of trees used in the forest | UniformInt (50, 250) |
| max_features | Number of features to observe when determining next split | UniformInt (2, n_features) | |
| max_depth | Maximum depth of the tree | UniformInt (1, 25) | |
| min_samples_split | Minimum number of samples required to split an internal node | UniformInt (2, 10) | |
| min_samples_leaf | The minimum number of samples required to be at a leaf node | UniformInt (1, 5) | |
| XGBoost | n_estimators | The number of boosting rounds | UniformInt (50, 250) |
| max_depth | The maximum depth of a tree | UniformInt (2, 10) | |
| subsample | The fraction of samples to be used for fitting the individual base learners | UniformFloat (0.5, 1.0) | |
| learning_rate | The learning rate (shrinkage) | UniformFloat (0.1, 0.5) | |
| colsample_bytree | The subsample ratio of columns when constructing each tree | UniformFloat (0.5, 1.0) | |
| colsample_bylevel | The subsample ratio of columns for each level | UniformFloat (0.5, 1.0) | |
| min_child_weight | The minimum sum of instance weight (hessian) needed in a child | UniformInt (1, 10) | |
| gamma | The minimum loss reduction required to make a further partition on a leaf node | UniformFloat (0, 1.0) | |
| reg_lambda | L2 regularization term on weights | Exp(UniformFloat (1e-2, 1e2)) | |
| reg_alpha | L1 regularization term on weights | Exp(UniformFloat (1e-2, 1e2)) | |
| ENet | C | Inverse of regularization strength, smaller values specify stronger regularization | Exp(UniformFloat (1e-2, 1e2)) |
| L1_ratio | The ElasticNet mixing parameter, with 0≤L1_ratio ≤1. For L1_ratio=0 the penalty is an L2 penalty. For L1_ratio=1 it is an L1 penalty | UniformFloat (0, 1) | |
| Linear SVC | C | Regularization parameter. The strength of the regularization is inversely proportional to C | Exp(UniformFloat (1e-2, 1e2)) |
The ASA classification was used in the binary logistic regression to set a benchmark for performance. The mode of each training set partition was used to handle missing values.
Model analysis
Performance of the final models was assessed on the hold-out testing set using Area Under the Receiver Operating Characteristic curve (AUROC) as the primary metric. The model with the highest mean AUROC and corresponding confidence intervals (based on standard deviation) was selected for further analysis. Sensitivity and specificity metrics were also determined for this model. To assess feature importance, we applied tree-dependent Shapley values using the SHAP package in Python on the final trained model, highlighting the ten most influential features for cervical fusion and decompression procedures [ Figure 1 ].
RESULTS
We included 4480 unique anterior and posterior cervical decompression and fusion surgeries in our analysis. On average, a patient in the cohort was 56.5 years old (SD 12.6), had a BMI of 29.4 (SD 6.21), and 47.8% were female. Most variables did not have missing data [ Supplementary Table 4 ]. Within the fusion cohort, the 75 th percentile thresholds for office visits, specialist referrals, ED/UC visits, pain medication prescriptions, and spine imaging orders were 3, 2, 0, 7, and 6, respectively. Of the cervical fusion surgeries, 12% of cases were identified as high health utilization.
Supplemental Table 4
Summary statistics and missing data
SD - Standard deviation, BMI - Body mass index, NSAIDs - Nonsteroidal anti-inflammatory drugs
| Variable | Cervical fusion summary statistics | |||
| Continuous | Mean±SD | Missing count | Missing percent | |
| BMI | 29.42±6.21 | 16 | 0.36 | |
| Age | 56.48±12.63 | 0 | 0 | |
| OR length | 236.64±107.27 | 0 | 0 | |
| Categorical | True count | Percent true | Missing count | Missing percent |
| Female | 2142 | 47.81 | 0 | 0 |
| AIDS and HIV | 27 | 0.6 | 0 | 0 |
| Lymphoma | 45 | 1 | 0 | 0 |
| Paralysis | 5 | 0.11 | 0 | 0 |
| Psychoses | 18 | 0.4 | 0 | 0 |
| Depression | 997 | 22.25 | 0 | 0 |
| Drug abuse | 93 | 2.08 | 0 | 0 |
| Weight loss | 55 | 1.23 | 0 | 0 |
| Coagulopathy | 129 | 2.88 | 0 | 0 |
| Alcohol abuse | 140 | 3.12 | 0 | 0 |
| Liver disease | 223 | 4.98 | 0 | 0 |
| Renal failure | 222 | 4.96 | 0 | 0 |
| Hypothyroidism | 524 | 11.7 | 0 | 0 |
| Valvular disease | 212 | 4.73 | 0 | 0 |
| Blood loss anemia | 21 | 0.47 | 0 | 0 |
| Deficiency anemia | 186 | 4.15 | 0 | 0 |
| Metastatic cancer | 273 | 6.09 | 0 | 0 |
| Cardiac arrhythmia | 465 | 10.38 | 0 | 0 |
| Rheumatoid arthritis | 319 | 7.12 | 0 | 0 |
| Diabetes complicated | 203 | 4.53 | 0 | 0 |
| Diabetes uncomplicated | 746 | 16.65 | 0 | 0 |
| Congestive heart failure | 167 | 3.73 | 0 | 0 |
| Hypertension complicated | 27 | 0.6 | 0 | 0 |
| Chronic pulmonary disease | 814 | 18.17 | 0 | 0 |
| Solid tumor without metastasis | 398 | 8.88 | 0 | 0 |
| Hypertension uncomplicated | 2074 | 46.29 | 0 | 0 |
| Other neurological disorder | 231 | 5.16 | 0 | 0 |
| Peripheral vascular disorder | 192 | 4.29 | 0 | 0 |
| Pulmonary circulation disorder | 127 | 2.83 | 0 | 0 |
| Fluid and electrolyte disorders | 135 | 3.01 | 0 | 0 |
| Peptic ulcer disease excluding bleeding | 140 | 3.12 | 0 | 0 |
| Validation cur tobacco | 282 | 6.29 | 0 | 0 |
| Validation fmr tobacco | 336 | 7.5 | 0 | 0 |
| Shx cervical fusion | 1428 | 31.87 | 0 | 0 |
| Shx cervical surgery | 837 | 18.68 | 0 | 0 |
| Shx thoracolumbar fusion | 228 | 5.09 | 0 | 0 |
| Shx thoracolumbar surgery | 393 | 8.77 | 0 | 0 |
| Shx unspecified spine fusion | 197 | 4.4 | 0 | 0 |
| Shx unspecified spine surgery | 385 | 8.59 | 0 | 0 |
| Cpt acdf | 3539 | 79 | 0 | 0 |
| Cpt pcdf c2below | 1036 | 23.12 | 0 | 0 |
| Cpt pcdf occiput c2 | 54 | 1.21 | 0 | 0 |
| Cpt pcdf c1c2 | 36 | 0.8 | 0 | 0 |
| Lumbar stenosis | 374 | 8.35 | 0 | 0 |
| Lumbar spondylosis | 53 | 1.18 | 0 | 0 |
| Lumbar disc disorders | 296 | 6.61 | 0 | 0 |
| Cervical disc disorders | 1524 | 34.02 | 0 | 0 |
| Cervical stenosis | 1470 | 32.81 | 0 | 0 |
| Cervical spondylosis | 17 | 0.38 | 0 | 0 |
| Cervical disc herniation | 989 | 22.08 | 0 | 0 |
| Lumbar disc herniation | 151 | 3.37 | 0 | 0 |
| Spine preop pseudoarthrosis post fusion | 13 | 0.29 | 0 | 0 |
| Spine preop post laminectomy syndrome | 40 | 0.89 | 0 | 0 |
| ACE inhibitor | 406 | 9.06 | 0 | 0 |
| ARB | 319 | 7.12 | 0 | 0 |
| Antidepressant | 837 | 18.68 | 0 | 0 |
| Beta 2 agonist | 719 | 16.05 | 0 | 0 |
| Beta blocker | 2065 | 46.09 | 0 | 0 |
| Benzodiazepine | 1947 | 43.46 | 0 | 0 |
| Immunosuppressant | 2580 | 57.59 | 0 | 0 |
| NSAID | 1065 | 23.77 | 0 | 0 |
| Opioid | 3055 | 68.19 | 0 | 0 |
| Anti-psychotic | 1977 | 44.13 | 0 | 0 |
| Neuromodulator | 1139 | 25.42 | 0 | 0 |
| Bisphosphonate | 7 | 0.16 | 0 | 0 |
| Loop diuretic | 128 | 2.86 | 0 | 0 |
| Thiazide diuretic | 264 | 5.89 | 0 | 0 |
| Cinacalcet | 1 | 0.02 | 0 | 0 |
| Insulin | 245 | 5.47 | 0 | 0 |
| Oral diabetes | 227 | 5.07 | 0 | 0 |
| Calcium supplement | 993 | 22.17 | 0 | 0 |
| Vitamin D supplement | 270 | 6.03 | 0 | 0 |
| Multilevel procedure | 2993 | 66.81 | 0 | 0 |
| Instrumentation | 4170 | 93.08 | 0 | 0 |
| General anesthesia use | 4355 | 97.82 | 28 | 0.62 |
| Inpatient procedure | 4107 | 91.69 | 1 | 0.02 |
| High health utilizer | 552 | 12.32 | 0 | 0 |
Cervical fusion health utilization outcomes were best predicted by the BRF model, which achieved an AUROC of 0.772 ± 0.007, significantly higher than the ASA benchmark model of 0.614 ± 0.007 [ Table 1 ]. Of the top-ten most important features determined by SHAP, 8 of the 10 were positively associated with high healthcare utilization. These variables include OR Duration, preoperative neuromodulator use at 90 days, preoperative opioid use at 90 days, preoperative immunosuppressant use at 90 days, preoperative benzodiazepine use at 90 days, C2 and below (subaxial) posterior cervical decompression and fusion (PCDF) procedure, preoperative beta blocker use at 90 days, and a medical history of depression. Anterior cervical decompression and fusion (ACDF) procedures were indicated to be protective of high health utilization, and age at the time of procedure was deemed insignificant.
Table 1
Machine learning algorithm performance for anterior and posterior cervical decompression and fusion cohort
ASA - American society of anesthesiologists; NN - Neural network; AUC - Area under the curve; CI - confidence interval; RF - Random forest; ENet - Elastic net; XGB - Extreme gradient boosted tree; SVC - Support vector classification
| Model | Fusion AUC (CI) |
| RF | 0.772 (0.007) |
| ENet | 0.77 (0.008) |
| XGB | 0.766 (0.007) |
| Linear SVC | 0.763 (0.008) |
| NN | 0.756 (0.008) |
| ASA | 0.614 (0.008) |
DISCUSSION
To the best of our knowledge, the presented retrospective study is the first to apply ML techniques to predict high outpatient healthcare utilization after anterior or posterior cervical spine surgery. Several tested models outperformed the ASA clinical benchmark prediction; of note, the balanced random forest classifier demonstrated the highest performance, with an Area under the curve of 0.772, sensitivity of 0.646, and specificity of 0.752, and therefore was a significantly more balanced predictor of utilization than the logistic regression baseline. SHAP analysis revealed the most pertinent features for prediction, including time spent in the OR, preoperative neuromodulator use, opioid use, and benzodiazepine exposure, followed by beta blocker and immunosuppressant use and age. Taken together, our models provide an objective gauge of the most important determinants of outpatient healthcare utilization with a considerable degree of confidence and discrimination relative to the current gold standard. Importantly, these findings suggest that even in the absence of surgical variables or postoperative data, patients at elevated risk for ED revisits and need for advanced imaging, pain medication refills, and specialist refills within 90 days of surgery can be effectively identified using such approaches.
As the volume of clinical data available and our understanding of multifactorial influences on outcomes following spine surgery grow, ML models are commonly deployed to elucidate the predictive significance of individual patient care attributes. Arvind et al . highlighted that ML models that incorporated comorbidities derived from the EHR outperformed the ASA classification in predicting adverse events following ACDF.[ 13 ] Similarly, Han et al . found that nonclinical factors, such as Medicaid status and ED visit history, are also significantly prognostic.[ 14 ] Our findings further underline the importance of such holistic considerations, moving beyond analyses emphasizing readmission alone, and address a dimension that current studies do not heavily value.
Our findings successfully capture several key predictors.[ 4 ] Across all tested models, preoperative prescriptions – particularly those associated with psychiatric diagnoses and chronic illnesses – demonstrated high feature importance. Preoperative opioid, neuromodulator, and benzodiazepine use were among the most consistently significant predictors of high postoperative healthcare utilization. These findings are consistent with the literature; preoperative opioid use is a strong predictor of continued postoperative opioid dependence, worse pain control, and higher complication rates, particularly in cervical spine surgery patients.[ 15 , 16 , 17 ] Kalakoti et al . demonstrated that among cervical fusion patients, chronic opioid therapy is associated with significantly higher odds of 90-day ED visits and reoperation rates.[ 18 ] A similar study corroborated these findings, further highlighting that preoperative opioid use is associated with continued postoperative dependence and may compound rates of postoperative healthcare resource access.[ 19 ] Benzodiazepine and neuromodulator use also often reflects underlying psychiatric comorbidities, chronic pain syndromes (particularly neuropathic), and psychosocial needs, which may contribute to complex recovery trajectories and are not well captured by traditional surgical risk scoring systems. At baseline, such patients may experience amplified perceived pain, stifled functional recovery, and difficulties with medication withdrawal; with the addition of surgical intervention, this likely leads to more frequent ED visits or uncontrolled symptoms requiring extensive follow-up.[ 20 , 21 ] However, more rigid indices such as the ASA or Charlson Comorbidity Index do not capture such factors in depth.[ 13 ] Medication-based indicators of chronic illness and mental health must be integrated for more accurate prediction of outpatient needs.
Preoperative use of beta blockers and immunosuppressants is similarly associated with systemic illness, such as cardiovascular disease and autoimmune disorders. These are established risk factors among spine surgery patients, predisposing individuals to perioperative instability, reduced physiologic reserve, and delayed mobilization secondary to poor wound healing and increased risk of infection.[ 22 , 23 ] Even when surgery-related complications are managed, these patient populations are known to require more prudent follow-up given greater baseline risk and burden.[ 23 , 24 ] Taken together, these results reinforce the narrative that postoperative outcomes are shaped less by single-point surgical decisions and more by a patient’s longitudinal engagement with the healthcare system. Incorporating such medication patterns into surgical planning and discharge workflows may enable earlier identification of at-risk patients and help mitigate downstream healthcare burden or fragmentation in transition of care.
We also flag traditional drivers of healthcare utilization in our findings. Longer operative times, age, and posterior fusions below C2 are expectedly associated with greater adverse predictions.[ 25 , 26 ] These are well-documented characteristics in the literature, as patients belonging to high-risk populations and those undergoing greater degrees of spinal intervention have a greater likelihood of needing more extensive postoperative support.[ 27 , 28 ] Our findings also support the commonly cited “protective” effects of an anterior versus posterior surgical approach. National registry studies have reported that PCDF is associated with greater patient mortality, higher complication rates, and longer hospital stays than ACDF, even after adjusting for case complexity.[ 29 , 30 , 31 ] PCDF is generally a more extensive procedure and is reserved for advanced pathology, likely contributing to the need for increased postoperative monitoring. Older age can amplify many of these challenges, as elderly patients have both reduced physiological resilience and a greater likelihood of presenting with comorbidities requiring concurrent follow-up postoperatively. These findings highlight the need for careful personalization of surgical paradigms to patients, especially those with vulnerabilities that may be exacerbated by intensive intervention.[ 32 ]
Our findings are particularly significant in underscoring how cervical spine surgery protocols can be shaped to better serve at-risk patient populations, moving beyond just considerations of risk of readmission. Several predictors in our cohort are modifiable within the bounds of preoperative care, and thus, management can be accordingly tailored if these factors are acknowledged in care planning. Preoperative use of opioids and benzodiazepines can be tapered if patients receive coordinated psychiatric care, as weaning protocols have been shown to improve outcomes.[ 33 ] Groups have employed a range of strategies among spine surgery patients, from tapering plans with complementary telephone counseling to preoperative educational programs to mitigate opioid use.[ 34 , 35 ] Similarly, increased neuromodulators should be recognized in patients, who may additionally benefit from multidisciplinary pain management workflows. These upstream changes may attenuate resource-intensive, delayed recoveries. Such efforts also stand to mitigate psychosocial stressors that may drive unnecessary healthcare encounters. Among patients with chronic conditions necessitating immunosuppressants or beta blockers, optimized medical clearance with comprehensive infection prophylaxis or tighter control of cardiovascular conditions may prevent the long-term utilization that our models capture. The standardization of these evidence-based protocols for groups of patients stands to bolster more efficient resource utilization, shorten hospital stays, and improve patient satisfaction. Future work should focus on integrating a broader range of patient socioeconomic-level data and multimodal data, such as patient imaging, to enhance predictive value.[ 36 ] The evolution of prognostic scales away from generics and instead toward patient-calibrated metrics serves to improve patient care on all fronts, contributing to the growing paradigm of proactive, personalized resource distribution.
Although the presented findings are significant, there are several limitations to the study. Our data is derived from a single institution and tertiary care center, limiting the generalizability of our findings to other varied care settings. Further, EHR data does not presently include patient reported outcomes to characterize true patient satisfaction or intraoperative details, particularly to characterize subjective drivers – such as perceived pain – of continued utilization. EHR limitations also extend to our definition of high utilization, which may not capture visits outside of our institutional network. Third, our follow-up window is limited to 90 days and may therefore not wholly capture a spectrum short- and long-term healthcare utilization, such as hardware issues or those pertaining to adjacent segment disease. Finally, while our models demonstrated balanced performance, class imbalance apparent across certain categories would be resolved by a larger dataset.
CONCLUSION
We believe that our study provides a modeling approach for significant, data-driven assessments of postoperative healthcare utilization among cervical spine surgery. With further validation and refinement of these tools, we strongly believe that these models can be a valuable tool for spine surgeons.
Conflicts of interest
There are no conflicts of interest.
Funding Statement
Nil.
REFERENCES
1. Philipp LR, Leibold A, Mahtabfar A, Montenegro TS, Gonzalez GA, Harrop JS. Achieving value in spine surgery: 10 major cost contributors. Global Spine J. 2021;11:14S–22S. doi: 10.1177/2192568220971288.
2. Goyal A, Bhandarkar AR, Alvi MA, Kerezoudis P, Yolcu YU, Habermann EB, et al. Cost of readmissions following anterior cervical discectomy and fusion: Insights from the nationwide readmissions database. Neurosurgery. 2020;87:679–88. doi: 10.1093/neuros/nyz443.
3. Gilmartin NF, Carreon LY, Gum JL, Dimar JR, 2nd, Brown ME, Glassman SD. Costs associated with potentially unnecessary postoperative healthcare encounters after lumbar spine surgery. Spine J. 2022;22:265–71. doi: 10.1016/j.spinee.2021.08.009.
4. Kammien AJ, Galivanche AR, Gouzoulis MJ, Moore HG, Mercier MR, Grauer JN. Emergency department visits within 90 days of single-level anterior cervical discectomy and fusion. N Am Spine Soc J. 2022;10:100122. doi: 10.1016/j.xnsj.2022.100122.
5. Liu CY, Zygourakis CC, Yoon S, Kliot T, Moriates C, Ratliff J, et al. Trends in utilization and cost of cervical spine surgery using the national inpatient sample database, 2001 to 2013. Spine (Phila Pa 1976) 2017;42:E906–13. doi: 10.1097/BRS.0000000000001999.
6. Oglesby M, Fineberg SJ, Patel AA, Pelton MA, Singh K. Epidemiological trends in cervical spine surgery for degenerative diseases between 2002 and 2009. Spine (Phila Pa 1976) 2013;38:1226–32. doi: 10.1097/BRS.0b013e31828be75d.
7. Kuo CY, Yu LC, Chen HC, Chan CL. Comparison of models for the prediction of medical costs of spinal fusion in Taiwan diagnosis-related groups by machine learning algorithms. Healthc Inform Res. 2018;24:29–37. doi: 10.4258/hir.2018.24.1.29.
8. Gowd AK, O’Neill CN, Barghi A, O’Gara TJ, Carmouche JJ. Feasibility of machine learning in the prediction of short-term outcomes following anterior cervical discectomy and fusion. World Neurosurg. 2022;168:e223–32. doi: 10.1016/j.wneu.2022.09.090.
9. Karabacak M, Margetis K. Development of personalized machine learning-based prediction models for short-term postoperative outcomes in patients undergoing cervical laminoplasty. Eur Spine J. 2023;32:3857–67. doi: 10.1007/s00586-023-07923-x.
10. Karabacak M, Bhimani AD, Schupper AJ, Carr MT, Steinberger J, Margetis K. Machine learning models on a web application to predict short-term postoperative outcomes following anterior cervical discectomy and fusion. BMC Musculoskelet Disord. 2024;25:401. doi: 10.1186/s12891-024-07528-5.
11. Karabacak M, Margetis K. Machine learning-based prediction of short-term adverse postoperative outcomes in cervical disc arthroplasty patients. World Neurosurg. 2023;177:e226–38. doi: 10.1016/j.wneu.2023.06.025.
12. Khazanchi R, Bajaj A, Shah RM, Chen AR, Reyes SG, Kurapaty SS, et al. Using machine learning and deep learning algorithms to predict postoperative outcomes following anterior cervical discectomy and fusion. Clin Spine Surg. 2023;36:143–9. doi: 10.1097/BSD.0000000000001443.
13. Arvind V, Kim JS, Oermann EK, Kaji D, Cho SK. Predicting surgical complications in adult patients undergoing anterior cervical discectomy and fusion using machine learning. Neurospine. 2018;15:329–37. doi: 10.14245/ns.1836248.124.
14. Han SS, Azad TD, Suarez PA, Ratliff JK. A machine learning approach for predictive models of adverse events following spine surgery. Spine J. 2019;19:1772–81. doi: 10.1016/j.spinee.2019.06.018.
15. Gerlach EB, Plantz MA, Swiatek PR, Wu SA, Arpey N, Fei-Zhang D, et al. The drivers of persistent opioid use and its impact on healthcare utilization after elective spine surgery. Global Spine J. 2024;14:370–9. doi: 10.1177/21925682221104731.
16. Lee D, Armaghani S, Archer KR, Bible J, Shau D, Kay H, et al. Preoperative opioid use as a predictor of adverse postoperative self-reported outcomes in patients undergoing spine surgery. J Bone Joint Surg Am. 2014;96:e89. doi: 10.2106/JBJS.M.00865.
17. Menendez ME, Ring D, Bateman BT. Preoperative opioid misuse is associated with increased morbidity and mortality after elective orthopaedic surgery. Clin Orthop Relat Res. 2015;473:2402–12. doi: 10.1007/s11999-015-4173-5.
18. Kalakoti P, Volkmar AJ, Bedard NA, Eisenberg JM, Hendrickson NR, Pugely AJ. Preoperative chronic opioid therapy negatively impacts long-term outcomes following cervical fusion surgery. Spine (Phila Pa 1976) 2019;44:1279–86. doi: 10.1097/BRS.0000000000003064.
19. Chung JH, Mierke A, Ramos O, Kagabo W, Boling W, Cheng W, et al. Preoperative opiate use leads to increased postoperative opiate use and readmissions after anterior cervical discectomy and fusion. J Spine Surg. 2022;8:214–23. doi: 10.21037/jss-21-126.
20. Horenstein A, Heimberg RG. Anxiety disorders and healthcare utilization: A systematic review. Clin Psychol Rev. 2020;81:101894. doi: 10.1016/j.cpr.2020.101894.
21. Romanelli RJ, Shah SN, Ikeda L, Lynch B, Craig TL, Cappelleri JC, et al. Patient characteristics and healthcare utilization of a chronic pain population within an integrated healthcare system. Am J Manag Care. 2017;23:e50–6.
22. Campbell PG, Yadla S, Nasser R, Malone J, Maltenfort MG, Ratliff JK. Patient comorbidity score predicting the incidence of perioperative complications: Assessing the impact of comorbidities on complications in spine surgery. J Neurosurg Spine. 2012;16:37–43. doi: 10.3171/2011.9.SPINE11283.
23. Badiee RK, Mayer R, Pennicooke B, Chou D, Mummaneni PV, Tan LA. Complications following posterior cervical decompression and fusion: A review of incidence, risk factors, and prevention strategies. J Spine Surg. 2020;6:323–33. doi: 10.21037/jss.2019.11.01.
24. Ahmad W, Fernandez L, Bell J, Krol O, Kummer N, Passfall L, et al. Assessment of postoperative outcomes of spine fusion patients with history of cardiac disease. J Am Acad Orthop Surg. 2022;30:e683–9. doi: 10.5435/JAAOS-D-21-00850.
25. Qi W, Mei Z, Mao X, Zhu L, Shao Y, Ge G, et al. Causal associations between sarcopenia-related traits and intervertebral disc degeneration: A two-sample Mendelian randomization analysis. Eur Spine J. 2024;33:2430–8. doi: 10.1007/s00586-024-08291-w.
26. Rajan PV, Emara AK, Ng M, Grits D, Pelle DW, Savage JW. Longer operative time associated with prolonged length of stay, non-home discharge and transfusion requirement after anterior cervical discectomy and fusion: An analysis of 24,593 cases. Spine J. 2021;21:1718–28. doi: 10.1016/j.spinee.2021.04.023.
27. Badhiwala JH, Ellenbogen Y, Khan O, Nouri A, Jiang F, Wilson JR, et al. Comparison of the inpatient complications and health care costs of anterior versus posterior cervical decompression and fusion in patients with multilevel degenerative cervical myelopathy: A retrospective propensity score-matched analysis. World Neurosurg. 2020;134:e112–9. doi: 10.1016/j.wneu.2019.09.132.
28. Wang D, Liao W, Hu H, Lei X, Zheng X, Jin D. Risk factors for ninety-day readmission following cervical surgery: A meta-analysis. J Orthop Surg Res. 2022;17:477. doi: 10.1186/s13018-022-03377-x.
29. Memtsoudis SG, Hughes A, Ma Y, Chiu YL, Sama AA, Girardi FP. Increased in-hospital complications after primary posterior versus primary anterior cervical fusion. Clin Orthop Relat Res. 2011;469:649–57. doi: 10.1007/s11999-010-1549-4.
30. Sattari SA, Ghanavatian M, Feghali J, Rincon-Torroella J, Yang W, Xu R, et al. Anterior cervical discectomy and fusion versus posterior decompression in patients with degenerative cervical myelopathy: A systematic review and meta-analysis. J Neurosurg Spine. 2023;38:631–43. doi: 10.3171/2023.1.SPINE221244.
31. Leckie S, Yoon ST, Isaacs R, Radcliff K, Fessler R, Haid R, Jr., et al. Perioperative complications of cervical spine surgery: Analysis of a prospectively gathered database through the association for collaborative spinal research. Global Spine J. 2016;6:640–9. doi: 10.1055/s-0035-1570089.
32. Khan O, Badhiwala JH, Witiw CD, Wilson JR, Fehlings MG. Machine learning algorithms for prediction of health-related quality-of-life after surgery for mild degenerative cervical myelopathy. Spine J. 2021;21:1659–69. doi: 10.1016/j.spinee.2020.02.003.
33. McAnally H. Rationale for and approach to preoperative opioid weaning: A preoperative optimization protocol. Perioper Med (Lond) 2017;6:19. doi: 10.1186/s13741-017-0079-y.
34. Uhrbrand PG, Rasmussen MM, Haroutounian S, Nikolajsen L. An individualised tapering protocol reduces opioid use 1 year after spine surgery: A randomised controlled trial of patients with preoperative opioid use. Acta Anaesthesiol Scand. 2023;67:1085–90. doi: 10.1111/aas.14266.
35. Joo SS, Hunter OO, Tamboli M, Leng JC, Harrison TK, Kassab K, et al. Implementation of a patient-specific tapering protocol at discharge decreases total opioid dose prescribed for 6 weeks after elective primary spine surgery. Reg Anesth Pain Med. 2020;45:474–8. doi: 10.1136/rapm-2020-101324.
36. Reyes SG, Bajaj PM, Herrera DE, Kurapaty SS, Chen A, Khazanchi R, et al. Predictive value of social determinants of health on 90-day readmission and health utilization following ACDF: A comparative analysis of XGBoost, random forest, elastic-net, SVR, and deep learning. Global Spine J. 2025;15:3797–806. doi: 10.1177/21925682251332556.