Outcome prediction is essential in the clinical decision-making process. By analyzing risk factors, you can predict the outcome of a determined event, such as in-hospital mortality after a medical operation.This can be used to decide whether to perform a particular operation depending on the outcome prediction.Artelnics worked with Hospital Rey Juan Carlos, located in Madrid, to predict in-hospital mortality following major lower extremity amputations (LEA) in type 2 diabetic patients using artificial intelligence to improve their clinical decision process.

Data set

In this project, we used the Spanish National Hospital Discharge Database (period 2003-2013) to select all hospital admissions of major LEA procedures in T2DM patients.

The data file analyzed included all procedures of major amputation, defined as any LEA through or proximal to the ankle joint. The final sample included 40,857 patients who had undergone major amputation procedures between January 1, 2003, and December 31, 2013.

Data analysis

Our team elaborated different neural network models to analyze which one offered the best results. The two models that offered the best result were:

  • All comordibities included in the Charlson comordibities index (CCI).
  • All comordibities included in the Elixhauser comordibities index (ECI).

Both models used a standard feed-forward, back-propagation neural network, each with its inputs, trained using the quasi-Newton training algorithm and optimized using model selection algorithms.

To comparate the models, a comparison of performance in terms of binary classification tests were made:

Indices Description Charlson Elixhauser
Accuracy rate Ratio of instances correctly classified 0.830 0.861
Error rate Ratio of instances misclassified 0.169 0.138
Sensitivity Portion of actual positives which are predicted positive 0.744 0.763
Specifity Portion of actual negative predicted negative 0.912 0.960

To have a better knowledge about which model to use, we studied the ROC curves.

ROC curve CCI model

The ECI model (that you can see above) showed a better area under the ROC curve 91.7% compared with the CCI model (91.7% [95% CI 90.3–93.0] vs 88.9% [95% CI; 87.5–90.2]), which means that it will correctly predict more cases.

After knowing that the CCI offered better results, Artelnics studied the different variable sensitive ratios (VSR) to know which variables are the best predictors for IHM following major ELA.

Variable ranking CCI variables Variable sensitive ratios
1st Age 1.577
2nd Female 1.559
3rd Myocardial infarction 1.477
4th Renal disease 1.456
5th Congestive heart failure 1.447
6th Moderate or severe liver disease 1.412
7th Metastatic solid tumor 1.362

Neural Network

Now that we have determined that the model to use is the Charlson comorbidity index, we can design a neural network able to predict in-hospital mortality in the studied cases. The inputs of the neural network corresponding to the different variables the CCI considers. Then, the two hidden layers will make the necessary calculations to predict the output, mortality.

Mortality neural network


With our work, we compared the two most commonly used comorbidity risk adjustment models, CCI and ECI, to determine which one adjusts better to type 2 diabetes mortality following major lower extremity amputation, underlying that some of the patient conditions make them more likely than others to have more health complications.

It can be seen from the study that the Charlson comorbidity offers slightly better results than can result in fewer errors.

IHM after major LEA in type 2 diabetes patients ranges from 7% to 12.4%. A further investigation into the motives behind the amputation decision would explain the difference in postoperative mortality rates.

One of the variables that affect the most is age. Older age is associated with a higher prevalence of comorbid conditions, which are also associated with higher IHM rates. It can also be seen that women have higher mortality rates in type 2 diabetes. In this graph, you can see the neural network created with the Charlson comorbidity index variables.

In conclusion, the predictors analyzed in our study could and should be applied in the clinical decision making to reduce mortality risk in type 2 diabetes patients.