Urologic Oncology: Seminars and Original Investigations
Clinical-Bladder cancerDynamic readmission prediction using routine postoperative laboratory results after radical cystectomy
Introduction
Radical cystectomy has one of the highest rates of complications and readmissions of any surgical procedure, with 25% of patients experiencing unplanned readmission within 30 days [1], [2], [3], [4]. These high readmission rates, coupled with increasing policy focus on reducing readmissions, have motivated investigations into identification and optimization of patients at highest readmission risk. However, the ability to predict readmission using traditional administrative data is limited making it unclear where and when to focus resources, leaving readmission rates largely unchanged [5,6].
There is increasing interest in incorporating dynamic data sources into readmission prediction models to better enable identification of high-risk cohorts [7]. While traditional administrative data are typically limited to static factors (eg, demographics, comorbidities), widespread use of electronic health records has made dynamic sources of data, laboratory results for example, readily available. The degree to which readily available laboratory data used to guide day-to-day clinical decision-making might impact readmission risk prediction after cystectomy is unknown. Indeed, such variables can be successfully incorporated into prediction models to improve performance for other outcomes ranging from transfer to the intensive care unit to mortality [8], [9], [10], [11]. For cystectomy patients with frequent postoperative lab draws, models using dynamic laboratory data could allow for better risk stratification and postoperative planning.
In this context, we used data from our institutional electronic health record to examine whether incorporating dynamic laboratory data into readmission prediction models improved risk stratification after radical cystectomy. Specifically, we assessed daily postoperative values for commonly obtained laboratory tests, and used machine learning techniques to compare values between readmitted and nonreadmitted patients. This study as a proof of concept, exploratory work demonstrates the unique promise of readily available, dynamic data to inform risk stratification of patients most likely to be readmitted after cystectomy.
Section snippets
Data source
We used an institutional database containing records on all inpatient and outpatient visits at our tertiary care facilities. This dataset was queried for all inpatient encounters associated with a diagnosis of bladder cancer (International Classification of Diseases 9th Revision code 188.X) and procedural codes for radical cystectomy (57.71) for the period from 2006 to 2016. This yielded a cohort of 996 patients who underwent radical cystectomy during the study period.
Outcomes and covariates
Our primary outcome for
Results
Among the 996 patients included in this cohort, 259 (26%) were readmitted within 30 days of discharge. Readmitted and nonreadmitted patients were similar in their demographic and clinical characteristics, though readmitted patients had higher BMI values, on average (Table 1, P < 0.01). Most patients were older, married, Caucasian men, and the minority were treated with robotic cystectomy.
As illustrated in Fig. 1, several of the laboratory tests in this study showed differences between
Discussion
Using machine learning techniques, we found differences in common postoperative laboratory values between readmitted and nonreadmitted patients treated with cystectomy. We also calculated threshold values to help differentiate patients at high and low risk of readmission within 30 days of discharge. Moreover, incorporating daily postoperative laboratory value thresholds into our readmission prediction models increased discriminatory power as measured by the c-statistic, when compared to models
Conclusions
We found that readmission risk assessment following radical cystectomy is significantly improved by the addition of dynamic physiologic data collected in modern electronic health record systems. Future work should refine these algorithms and study their implementation into daily practice in order to help guide clinical decision-making. While the problem of readmission after radical cystectomy appears refractory, innovative, dynamic approaches to existing data sources appear poised to enable
Conflicts of Interest
None.
References (22)
- et al.
No differences in population-based readmissions after open and robotic-assisted radical cystectomy: implications for post-discharge care
Urology
(2017) - et al.
Early and late perioperative outcomes following radical cystectomy: 90-day readmissions, morbidity and mortality in a contemporary series
J Urol
(2010) - et al.
Understanding hospital readmission intensity after radical cystectomy
J Urol
(2015) - et al.
Root causes and modifiability of 30-day hospital readmissions after radical cystectomy for bladder cancer
J Urol
(2016) - et al.
Development and validation of an electronic medical record-based alert score for detection of inpatient deterioration outside the ICU
J Biomed Inform
(2016) - et al.
Complications and failure to rescue after laparoscopic versus open radical nephrectomy
J Urol
(2011) - et al.
Sharpening the focus on causes and timing of readmission after radical cystectomy for bladder cancer
Cancer
(2014) - et al.
Few modifiable factors predict readmission following radical cystectomy
Can Urol Assoc J
(2015) - et al.
Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review
J Am Med Inform Assoc
(2017) - et al.
Risk-adjusting hospital inpatient mortality using automated inpatient, outpatient, and laboratory databases
Med Care
(2008)
Nonelective rehospitalizations and postdischarge mortality
Med Care
Cited by (2)
Artificial Intelligence Applications in Urology: Reporting Standards to Achieve Fluency for Urologists
2022, Urologic Clinics of North AmericaCitation Excerpt :Hasnain and colleagues24 evaluated a database with 3499 patients who had undergone radical cystectomy with a random forest model and was able to predict recurrence at 1 year and survival at years 1, 3, and 5, with a greater than 70% sensitivity and specificity. Kirk and colleagues25 trained 2 AI models using a support vector machine for feature selection, with logistic regression and random forest models to predict readmission in patients with bladder cancer with an AUC of 0.62 and 0.68, respectively. There were 2 publications identified that predicted outcomes in kidney cancer in our review of the literature (Table 11).