Access Type

Open Access Dissertation

Date of Award

January 2022

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Industrial and Manufacturing Engineering

First Advisor

Alper Murat

Abstract

Recently, health care related studies are being widely conducted by researchers using unique and efficient techniques to increase system profitability, quality of care, and patient satisfaction. Surgery department is considered as the hospital's engine, and cost of surgical services has a huge impact on the overall profitability of the hospital. This thesis proposes novel approaches to improve the efficiency of surgical services by using machine learning concepts.

In the first part, this research investigates the prediction of the surgery durations and Current Procedural Terminology (CPT) Codes. Accurate prediction of the surgery duration will improve the utilization of indispensable surgical resources such as surgeons, nurses, and operating rooms. Prediction of the correct CPT codes not only aids the preparation for the survery (i.e., case cart) but also enhances prediction of surgery duration distributions.

In predicting the CPT code(s) of each surgery, we use continuous, categorical and textual preoperative information as the independent features. Since information-rich textual information available perioperatively is mostly entered manually and thus is non-standardized (i.e. abbreviations) and prone to typos. Accordingly, direct usage of the raw text features leads to loss of text feature information. Thus, we first find the most informative text features from unstructured principal procedure and some physician notes through a novel text mining method for the detection and clustering of typos and abbreviations and efficiently reduces feature dimensionality. The output is a well-established in terms of typo correction and abbreviation detection and provides accuracy improvements in the prediction of CPTs as well as surgery durations. To predict CPTs, we first focus on the primary CPT prediction and evaluate the predictive performances of different filtering and set-based prediction strategies. While the primary CPT code is the most important determinant of surgery durations and perioperative planning tasks, surgeries often entail multiple procedures (i.e., auxiliary CPTs) which can greatly influence the surgery durations. Hence, by using multi-task learning concepts, we develop models to predict multiple CPT codes, i.e. set containing the CPTs of all procedures being performed in the operation.

For the surgery duration prediction, we compare direct methods (i.e., regression based prediction using all feature information) with two-step approach where we first predict primary or set-CPTs of the surgery and then, given the predicted CPT codes, we estimate a duration distribution for each surgery case. By first predicting the CPT, the two-step approach provides valuable planning information to the preoperative services in addition to the improvements in surgery duration predictions. We evaluate the improvements in surgery duration estimation by comparing direct approach versus two-step approach and primary versus set-CPT predictions. Whereas direct approach primarily estimates the mean duration, the two-step approach naturally leads to a distribution information. We also evaluate the distributional information quality of the two-step approach with those that can be elicited from the direct approaches. Lastly, two-step approach also allows for more specific prediction and operational planning of surgical service operations such as case scheduling.

In order to account for the duration estimation loss in the single CPT prediction approach, we modified the CPT selection by applying the Genetic optimization algorithm. GA enables us to select the optimal trees with respect to the two goals, predicting correct CPT or estimating more accurate duration, in the model's boosting step. Lastly, we can compare the duration output of the revised single CPT approach with the aforesaid approaches. The hospital may choose to produce more accurate durations or weigh more on CPT prediction for the surgery cases given the package of CPT / duration prediction tool.

Download

Share

COinS