Enables personalized perioperative risk prediction using machine learning models based on preoperative data

We used extreme gradient boosting, a machine learning technique, to create a model capable of predicting postoperative in-hospital mortality in individual patients with high accuracy. The most important variables were preoperative ordered red blood cell count, c-reactive protein, and age. Tabular data contributed the most to the model’s predictive value, while unstructured data such as free text had less impact on the model’s performance. The effects of individual risk factors and changes in individual factors were calculated and displayed graphically to allow model interpretation. In clinical routine, models may help physicians and patients to support informed decision-making.

As for the most important variables in our model, the number of red packed cells provided depends mainly on the type of procedure. The patient’s existing condition may also play a role. Most hospitals have standards for determining the number of units of blood to be donated before surgery and are based on valid guidelines and hospital transfusion standards. Therefore, this factor does not reflect the purely subjective assessment of physicians.

The relationship between age and mortality risk is not controversial. Age is factored into many conventional scores, such as his POSPOM and the Charlson Comorbidity Index, mentioned above.^6,20High C-reactive protein levels can indicate infection and are associated with poor cardiopulmonary function.High preoperative C-reactive protein levels have been shown to be associated with postoperative complications^{twenty one}These factors therefore appear to be associated with postoperative outcomes, supporting the validity of the model.

Preoperative risk assessment is essential to identify patients at high risk of morbidity and mortality and to develop perioperative strategies to minimize those risks.²Additionally, knowing the risks can help keep patients properly informed and involved in decisions about planned surgery. Therefore, current guidelines recommend using various scores to assess the risk of perioperative complications in patients.²However, the most common ASA-PS, with a recently reported AUROC of approximately 0.63, had only poor predictive power for postoperative mortality.^{twenty two} It therefore seemed unsuitable for reliable mortality prediction. Recently, more complex scores such as POSPOM and CCI (Charlson Comorbidity Index) are preferred.^6,20However, AUROC reported 0.64 on CCI and 0.65 on Modified Frailty Index, so the predictive power of these scores does not actually exceed that of ASA-PS^{twenty three}The original report of the POSPOM score by LeManach et al. It looked like he could catch up with the ASA showing his AUROC of 0.944, but it should be taken into account that verification of his POSPOM in each country, including surgical code matching, needs to be performed. German validation of his POSPOM reported by Layer et al. Shows only AUROC of 0.771^{twenty four}.

The potential for applying machine learning techniques to perioperative care was confirmed by a recent systematic review. This review noted that many models reached AUROC greater than 0.9, thus outperforming most conventional scores.^{twenty five}The review further confirmed that random forest and gradient boosting were used most frequently and gave the best model performance.⁸Our findings were consistent with these results as we achieved an AUROC of 0.95 and an acceptable precision-recall trade-off using XGBoost.²⁶.

However, effectively improving patient outcomes and adapting perioperative approaches to patients requires more than risk knowledge. Simplified conventional scores only provide population-level predictions. However, our machine learning model can identify modifiable risk factors for every patient, thus opening the door to personalized medicine in the field of anesthesia.

To help clinicians interpret such complex models, the impact of all parameters on the model’s predictive power can be calculated. In addition, we can calculate the change in risk when changing a single factor, provided that everything else remains the same. However, this is a very theoretical approach, with many factors interacting and it is not always known which factors are interrelated and how. As with traditional retrospective analyses, unidentified confounders and multicollinearity effects must be allowed.Although not important for model fit, multicollinearity most likely dilutes the impact of important factors²⁷Nevertheless, in selected cases the model is very useful in explaining the effects of preoperatively initiated optimization measures. However, in order to do this, we need to know the dependent factors and be able to mathematically model their interrelationships. Here are his two cases on two current topics as examples. It is an attempt to improve the patient’s blood management and the functional capacity of the patient preoperatively.

In the first example, we find that perioperative mortality risk decreases significantly with increasing hemoglobin levels. Current guidelines recommend a preoperative assessment of anemia and treatment with threshold hemoglobin values ​​below her 13 g/dL in men and below her 12 g/dL in menstruating women.²⁸It is worth noting that in our typical patient, after reaching a minimum value, the risk of death rises again with increasing Hb, although the optimal value was defined by the World Health Organization (WHO). It means that it is within the threshold.

Our second example fits the increasingly important problem of prehabilitation, especially for elderly and frail patients to build reduced reserves. Recently, a meta-analysis showed that preoperative optimization measures can reduce postoperative morbidity. Unfortunately, harmonized protocols and procedures do not yet exist, and the impact of prehabilitation on postoperative outcomes is still unknown. Since it is not yet clear which preoperative interventions will benefit which patients, multimodal concepts are mainly pursued.^29,30Here we demonstrate the effect of weight gain in a patient with cachexia and we can conclude that this patient would benefit from preoperative nutritional therapy. The effect of the increase is not very pronounced.

In addition to the aforementioned multicollinearity effects, we have to face some challenges and limitations. Here we only offer single-center studies with a significant number of patients. surgical procedures classified by procedure code). The German OPS describes surgical procedures at a very detailed level. Most of these codes did not appear in the model due to their low frequencies. However, this classification is primarily used for billing purposes and cannot be grouped without loss of information. So I refrained from consolidating the code. In addition, we intentionally left some variables redundant, such as the glomerular filtration rate (GFR), which was calculated according to two different formulas. because they give different values ​​in different patient groups. Another limitation of our model is the lack of intraoperative information that could definitively alter mortality risk compared to the preoperative setting.It is clear that information such as the duration of surgery, intraoperative blood loss, or occurrence of adverse events can have a significant impact on the postoperative course.³¹However, our model provides reliable information for evaluating and counseling patients during pre-anesthetic visits prior to elective surgery.

Clinical documentation is often incomplete and may lack data on individual variables. We addressed this problem by adding, for each variable, an additional dichotomous function containing information about its availability. Interestingly, only two of these dichotomies remained in the final model: ‘bilirubin available’ and ‘main diagnostic available’ (see Table A2 in the Appendix).

Another important factor is the often poor quality of routinely collected preoperative data. This fact is reflected in the model, as numerical and tabular data contribute the most to the predictions. Data derived from free text are not included in the top 200 variables. Therefore, many of the factors displayed in the model, especially laboratory values, are only surrogate parameters for organic disease that are better accounted for in survey results or physician reports. However, important information about the patient’s preoperative status is usually collected as free text in clinical routine. Such unstructured information can be handled poorly by such complex models. There are currently two ways to remedy this fact. On the one hand, natural language processing methods can be refined and integrated into models.There is evidence that including such algorithms in models improves the quality of predictions³²However, there is a very heterogeneous set of algorithms available, some of which have not yet been externally validated.³³Another option is to avoid the extensive use of free text and force users to complete and structured input through the user interface. This will require a major redesign of most clinical documentation tools. In an era where interoperability between different medical document systems is becoming more and more important, obtaining structured information in a unified document architecture is a key prerequisite. It is hoped that uniform nomenclature and syntactic and semantic standards will make the use of data or the creation of predictive models much easier in the future, as well as scientific evaluation.

In conclusion, our study demonstrates that machine learning models can be created to predict the risk of postoperative in-hospital mortality with superior accuracy over conventional scores. Because this model can be used to determine risk factors at the individual level, it provides a suitable basis for informed consent in high-risk patients. In addition, we made the model interpretable by calculating the impact of changes in modifiable risk factors for selected cases. Our model is therefore suitable for identifying individualized mortality risk and assessing the effects of modifying risk factors in future studies.