Interpretable machine learning models for hospital readmission prediction: a two-step extracted regression tree approach.

BACKGROUND: Advanced machine learning models have received wide attention in assisting medical decision making due to the greater accuracy they can achieve. However, their limited interpretability imposes barriers for practitioners to adopt them. Recent advancements in interpretable machine learning tools allow us to look inside the black box of advanced prediction methods to extract interpretable models while maintaining similar prediction accuracy, but few studies have investigated the specific hospital readmission prediction problem with this spirit. METHODS: Our goal is to develop a machine-learning (ML) algorithm that can predict 30- and 90- day hospital readmissions as accurately as black box algorithms while providing medically interpretable insights into readmission risk factors. Leveraging a state-of-art interpretable ML model, we use a two-step Extracted Regression Tree approach to achieve this goal. In the first step, we train a black box prediction algorithm. In the second step, we extract a regression tree from the output of the black box algorithm that allows direct interpretation of medically relevant risk factors. We use data from a large teaching hospital in Asia to learn the ML model and verify our two-step approach. RESULTS: The two-step method can obtain similar prediction performance as the best black box model, such as Neural Networks, measured by three metrics: accuracy, the Area Under the Curve (AUC) and the Area Under the Precision-Recall Curve (AUPRC), while maintaining interpretability. Further, to examine whether the prediction results match the known medical insights (i.e., the model is truly interpretable and produces reasonable results), we show that key readmission risk factors extracted by the two-step approach are consistent with those found in the medical literature. CONCLUSIONS: The proposed two-step approach yields meaningful prediction results that are both accurate and interpretable. This study suggests a viable means to improve the trust of machine learning based models in clinical practice for predicting readmissions through the two-step approach.

Therapeutic efficacy of human adipose mesenchymal stem cells in Crohn's colon fibrosis is improved by IFN-γ and kynurenic acid priming through indoleamine 2,3-dioxygenase-1 signaling.

BACKGROUND: Inflammatory bowel diseases (IBD) are chronic relapsing-remitting inflammatory diseases of the gastrointestinal tract that are typically categorized into two subtypes: Crohn's disease (CD) and ulcerative colitis (UC). Although MSCs therapy has achieved encouraging outcomes in IBD therapy, objective responses are limited in colon fibrosis stenosis owing to the complicated microenvironment of CD and MSCs heterogeneity of quality. Here, we chose IFN-γ and kynurenic acid (KYNA) to overcome the low response and heterogeneity of human adipose-derived MSCs (hADSCs) to treat IBD and expand the therapeutic effects based on the excellent ability of IFN-γ and KYNA to promote indoleamine 2,3-dioxygenase-1 (IDO-1) signaling, providing a potential protocol to treat IBD and fibrosis disease. METHODS: hADSCs were isolated, cultured, and identified from human abdominal adipose tissue. The CD pathology-like acute colitis and chronic colon fibrosis rat model was induced by 2,4,6-trinitrobenzen sulfonic acid (TNBS). hADSCs were pretreated in vitro with IFN-γ and KYNA and then were transplanted intravenously at day 1 and 3 of TNBS administration in colitis along with at day 1, 15, and 29 of TNBS administration in chronic colonic fibrosis. Therapeutic efficacy was evaluated by body weights, disease activity index, pathological staining, real-time PCR, Western blot, and flow cytometry. For knockout of IDO-1, hADSCs were transfected with IDO-1-targeting small gRNA carried on a CRISPR-Cas9-lentivirus vector. RESULTS: hADSCs treated with IFN-γ and KYNA significantly upregulated the expression and secretion of IDO-1, which has effectively ameliorated CD pathology-like colitis injury and fibrosis. Notably, the ability of hADSCs with IDO-1 knockout to treat colitis was significantly impaired and diminished the protective effects of the primed hADSCs with IFN-γ and KYNA. CONCLUSION: Inflammatory cytokines IFN-γ- and KYNA-treated hADSCs more effectively alleviate TNBS-induced colitis and colonic fibrosis through an IDO-1-dependent manner. Primed hADSCs are a promising new strategy to improve the therapeutic efficacy of MSCs and worth further research.

Oncolytic adenovirus inhibits malignant ascites of advanced ovarian cancer by reprogramming the ascitic immune microenvironment.

Malignant ascites frequently occur in patients with advanced ovarian cancer at initial diagnosis, and in almost all cases of relapse, they are closely related to poor prognosis, chemoresistance, and metastasis. To date, effective management strategies have been limited. In this study, we aimed to investigate the effects of oncolytic adenovirus (OV) on malignant ascites in a mouse model of advanced ovarian cancer. The results suggested that OV conferred an effective ability to reduce ascites development and prolong overall survival. Further analysis of the ascitic immune microenvironment revealed that OV treatment promoted T cell infiltration, activation, and differentiation into the effector phenotype; reprogrammed macrophages toward the M1-like phenotype; and increased the ratios of both CD8+ T cells to CD4+ T cells and M1 to M2 macrophages. However, immunosuppressive factors such as PD-1, LAG-3, and Tregs emerged after treatment. Combination therapy including OV, CSF-1R inhibitor PLX3397, and anti-PD-1 remarkably delayed the progression of ascites, and combination therapy induced a greater extent of T cell infiltration, proliferation, and activation. This study provides experimental and theoretical evidence for oncolytic virus-based treatment of malignant ascites, which may further contribute to advanced ovarian cancer therapy.

Dexamethasone enhances the lung metastasis of breast cancer via a PI3K-SGK1-CTGF pathway.

Dexamethasone (Dex), as a pretreatment agent, is widely used to attenuate the side effects of chemotherapy in breast cancer treatment. However, whether and how Dex affects breast cancer metastasis remain to be furtherly understood. In this study, we established several mouse breast cancer metastatic models to study the effect of Dex in vitro and in vivo. Transwell, Western Blot and RNA interference were applied to study the molecular mechanism of Dex in promoting breast cancer cell migration. Meanwhile, the effect of Dex on lung metastasis of breast cancer in Dex combined with PTX chemotherapy was discussed. Our results confirmed that Dex could promote breast cancer cell metastasis both in vitro and in vivo. Mechanistic studies revealed that this pro-metastatic effect of Dex was mediated by the GR-PI3K-SGK1-CTGF pathway in tumor cells. Ligation of Dex and glucocorticoid receptor (GR) on tumor cells activated the PI3K signaling pathway and upregulated serum glucocorticoid-inducible kinase 1 (SGK1) expression, and then increased the expression of connective tissue growth factor (CTGF) through Nedd4l-Smad2. Moreover, Dex was the leading factor for lung metastasis in a standard regimen for breast cancer treatment with paclitaxel and Dex. Importantly, targeting SGK1 with the inhibitor GSK650394 remarkably reduced lung metastasis in this regimen. Our present data provide new insights into Dex-induced breast cancer metastasis and indicate that SGK1 could be a candidate target for the treatment of breast cancer metastasis.

Proportions of Surgical Patients Discharged Home the Same or the Next Day Are Sufficient Data to Assess Cases' Contributions to Hospital Occupancy.

Introduction When the hospital census is high, perioperative medical directors or operating room (OR) managers may need to consider postponing some surgical cases scheduled to be performed within the next three workdays. This scenario has arisen at hospitals in regions with large increases in admissions due to coronavirus disease 2019 (COVID-19). We compare summary measures for hospital length of stay (LOS) to guide the OR manager having to decide which cases may need to be postponed to ensure a sufficient reserve of available inpatient beds. Methods We studied the 1,201,815 ambulatory and 649,962 inpatient elective cases with a major therapeutic procedure performed during 2018 at all 412 non-federal hospitals in Florida. The data were sorted by the hospital, and then by procedure category. Statistical comparisons of LOS were made pairwise among all procedure categories with at least 100 cases at (the) each hospital, using the chi-square test (LOS ≤ 1 day versus LOS > 1 day), Student's t-test with unequal variances, and the Wilcoxon-Mann-Whitney test. The comparisons among the three tests then were repeated having sorted the data by procedure category and making statistical comparisons among all hospitals with at least 100 cases for the procedure category. Results Whether using a criterion for statistical significance of P < 0.05 or P < 0.01, and whether compared with Student's t-test with unequal variances or Wilcoxon-Mann-Whitney test, the chi-square test had greater odds (i.e., greater statistical power) to detect differences in LOS (all four with P < 0.0001 and all 95% lower confidence limits for odds ratios ≥ 3.00). The findings were consistent when the data, first sorted by procedure category and then by probability distributions of LOS, were compared between hospitals (all P < 0.0001 and the 95% lower confidence limits for odds ratio ≥ 3.72). Conclusions For purposes of comparing procedure categories pairwise at the same hospital, there was no loss of information by summarizing the probability distributions using single numbers, the percentages of cases among patients staying longer than overnight. This finding substantially simplifies the mathematics for constructing dashboards or summaries of OR information system data to help the perioperative OR manager or medical director decide which cases may need to be postponed, when the hospital census is high, to provide a sufficient reserve of inpatient hospital beds.

Forecasting the Probability That Each Surgical Case Will Either Be Ambulatory or the Patient Will Remain in the Hospital Overnight Versus Having a Length of Stay of Two or More Days.

When the hospital census is high, perioperative medical directors or operating room (OR) managers sometimes need to review with surgical departments as to which surgical cases scheduled to be performed within the next three days may need to be postponed. Although distributions of hospital length of stay (LOS) are highly skewed, a surprisingly effective summary measure is the percentage of patients previously undergoing the same category of procedure as that scheduled whose LOS was zero or one day. We evaluated how to forecast each hospital's percentage of cases with LOS of <2 days, segmented by category of surgical procedure. The large teaching hospital studied included several inpatient adult surgical suites, an ambulatory surgery center, and a pediatric surgical suite. We included 98,540 cases in a training dataset to predict 24,338 cases in a test dataset. For each category of procedure, we calculated the cumulative count of cases among quarters, from the most recent quarter, second most recent quarter, and so forth up to the quarter resulting in at least 800 cases. If every quarter combined had fewer than 800 cases for a given category of procedure, we included all cases for that category. For each combination of category and quarter, we used the cumulative counts of cases and cases with LOS of <2 days, excluding the current quarter. Then, for each category of procedure, and for each of the preceding quarters included for the category, we used the cumulative counts to calculate the asymptotic standard error (SE) for the proportion of cases with LOS of <2 days. If all preceding quarters combined provided a sample size such that the estimated SE for the proportion exceeded 1.25%, we included all preceding quarters. The observed absolute percentage error was 0.76% (SE: 0.12%). This error was nearly 100-fold smaller than the percentage of cases to which it would be used (i.e., 0.76% versus 73.1% with LOS of <2 days). The principal weakness of the forecasting methodology was a small bias caused by a progressive reduction in the overall LOS over time. However, this bias is unlikely to be important for predicting cases' LOS when the hospital census is high. When performing these time series calculations quarterly, a reasonable approach is to perform calculations of both case counts and SEs for each category of procedure. We recommend using the fewest historical quarters, starting with the most recent quarter, either with at least 800 cases or an estimated asymptotic SE for the estimated percentage no greater than 1.25%. Applying our methodology with local LOS data will allow OR managers to estimate the number of patients on the elective OR schedule each day who will be hospitalized for longer than overnight, facilitating communication and decision-making with surgical departments when census considerations constrain the ability to run a full surgical schedule.

The impact of opening dedicated clinics on disease transmission during an influenza pandemic.

Dedicated clinics can be established in an influenza pandemic to isolate people and potentially reduce opportunities for influenza transmission. However, their operation requires resources and their existence may attract the worried-well. In this study, we quantify the impact of opening dedicated influenza clinics during a pandemic based on an agent-based simulation model across a time-varying social network of households, workplaces, schools, community locations, and health facilities in the state of Georgia. We calculate performance measures, including peak prevalence and total attack rate, while accounting for clinic operations, including timing and location. We find that opening clinics can reduce disease spread and hospitalizations even when visited by the worried-well, open for limited weeks, or open in limited locations, and especially when the clinics are in operation during times of highest prevalence. Specifically, peak prevalence, total attack rate, and hospitalization reduced 0.07-0.32%, 0.40-1.51%, 0.02-0.09%, respectively, by operating clinics for the pandemic duration.

An Integrated Framework for Reducing Hospital Readmissions using Risk Trajectories Characterization and Discharge Timing Optimization.

When patients leave the hospital for lower levels of care, they experience a risk of adverse events on a daily basis. The advent of value-based purchasing among other major initiatives has led to an increasing emphasis on reducing the occurrences of these post-discharge adverse events. This has spurred the development of new prediction technologies to identify which patients are at risk for an adverse event as well as actions to mitigate those risks. Those actions include pre-discharge and post-discharge interventions to reduce risk. However, traditional prediction models have been developed to support only post-discharge actions; predicting risk of adverse events at the time of discharge only. In this paper we develop an integrated framework of risk prediction and discharge optimization that supports both types of interventions: discharge timing and post-discharge monitoring. Our method combines a kernel approach for capturing the non-linear relationship between length of stay and risk of an adverse event, with a Principle Component Analysis method that makes the resulting estimation tractable. We then demonstrate how this prediction model could be used to support both types of interventions by developing a simple and easily implementable discharge timing optimization.

Comparing Policies for Case Scheduling Within 1 Day of Surgery by Markov Chain Models.

BACKGROUND: In previous studies, hospitals' operating room (OR) schedules were influenced markedly by decisions made within a few days of surgery. At an academic hospital, 46% of ORs had their last case scheduled or changed within 1 working day of surgery, and a private hospital had 64%. Many of these changes were for patients who were admitted before surgery (i.e., inpatient cases). In this study, we investigate the impact on OR productivity of how cases are scheduled within 1 working day before the day of surgery. METHODS: We consider the case-scheduling choice between 2 ORs. We compare 3 scheduling policies: Best Fit Descending, Worst Fit Descending, and Worst Fit Ascending. "Descending" strategies consider new cases from longest to shortest, whereas "Ascending" considers new cases from shortest to longest. Best Fit schedules each new case into the OR with sufficient but the least remaining underutilized OR time for the case. Worst Fit does the same but with the most remaining time. For our application, Best Fit chooses a later start time, whereas Worst Fit chooses an earlier start time. In our computational model, cases are of 2 possible durations, brief or long. Case cancellation is incorporated explicitly, and the number of new cases to schedule depends on the current number of scheduled cases in each OR, both new from previous studies. The number of cases in each OR is modeled as a Markov chain, evolving between 2 periods, corresponding to 1 day and 0 days before the day of surgery. For each scheduling policy, we evaluate the mean overutilized OR time and productivity. Our sensitivity analyses cover many cancellation rates, arrival settings, case durations, and initial conditions (i.e., how cases are scheduled into the 2 ORs preceding 1 workday before the day of surgery). RESULTS: Best Fit Descending and Worst Fit Descending achieved almost the same overutilized time and productivity. Worst Fit Ascending caused greater overutilized time (as much as 6.6 minutes more per OR) and thus lesser productivity (as much as 1.6% less) compared with Best Fit Descending or Worst Fit Descending. When the staff were scheduled for less time than the optimal allocated OR time, there were nearly the same differences between the staff productivity resulting from the use of Worst Fit Ascending rather than Worst Fit Descending or Best Fit Descending. CONCLUSIONS: Scheduling office decision making within 1 day before surgery should be based on statistical forecasts of expected total OR workload (i.e., forecasts that include the addition of non-elective cases and the subtraction of cases that cancel). As long as a case is not scheduled into overutilized time when less overutilized time could be achieved in another OR, and cases are considered in descending sequence of scheduled durations, the differences in overutilized time and productivity among the scheduling policies are small. Cognitive bias in staff scheduling causes a significant reduction in productivity, but the differences among scheduling policies are nearly the same as when there is no bias.

Descriptive study of case scheduling and cancellations within 1 week of the day of surgery.

BACKGROUND: We performed a descriptive study of operating room (OR) case scheduling within 1 week of the day of surgery. METHODS: The data used were from the case scheduling and transaction audit tables of a hospital's anesthesia and OR information management systems. Each change to a scheduled case in the OR information system was captured in an audit table, including the date and time when the change was made. The timestamps allowed reconstruction of the elective OR schedule for each date of surgery at preceding dates (e.g., 2 workdays ahead). The sample size was n = 17 consecutive 4-week periods. The allocated OR time, for each combination of service and day of the week, was the number of hours that minimized the inefficiency of use of OR time, a weighted combination of the hours of underutilized OR time and the more expensive hours of overutilized OR time. Data are reported as mean ±SE. RESULTS: (1) The percentage of OR date combinations with at least 1 add-on case was 24.1% ± 0.3%. The most recent addition of a case to an OR occurred 1 working day before surgery for 22.3% ± 0.4% of OR date combinations. At least half (51.5% ± 0.5%) of ORs had its last case scheduled or changed within 2 working days of surgery. In addition, when allocated OR time was filled and the service scheduled additional case(s), the median time ahead when each such case was scheduled was 2.2 ± 0.2 workdays. Thus, managers can productively focus on the day of surgery starting 2 working days before surgery. (2) Once allocated time was full, the ratio of the net additional cases scheduled to the total number performed was 1.2% ± 0.6%. However, 11.1% ± 1.7% of the total were additional cases. Thus, schedulers should rely on the allocated time to be predictive of the actual (net) workload that will occur in the future, on the day of surgery. (3) For service and day combinations for which 2 working days ahead the scheduled hours exceeded the allocated hours, there was no significant net increase in minutes of cases scheduled (P = 0.79), unlike when the scheduled hours were less than allocated (P < 0.0001). Thus, additional hours of cases scheduled within the same number of workdays are heterogeneous both within and among services based on the prior hours of cases scheduled. CONCLUSIONS: Planning anesthesia assignments, ORs to target, etc., can be done productively starting 2 working days ahead of surgery. There are so many changes to the OR schedule those last 2 workdays that anesthesia groups should be engaged with the scheduling office during that period. The primary predictor of additional net hours of cases to be scheduled is the difference between the allocated (i.e., forecasted) OR time and the hours scheduled so far.

The impact of mass gatherings and holiday traveling on the course of an influenza pandemic: a computational model.

BACKGROUND: During the 2009 H1N1 influenza pandemic, concerns arose about the potential negative effects of mass public gatherings and travel on the course of the pandemic. Better understanding the potential effects of temporal changes in social mixing patterns could help public officials determine if and when to cancel large public gatherings or enforce regional travel restrictions, advisories, or surveillance during an epidemic. METHODS: We develop a computer simulation model using detailed data from the state of Georgia to explore how various changes in social mixing and contact patterns, representing mass gatherings and holiday traveling, may affect the course of an influenza pandemic. Various scenarios with different combinations of the length of the mass gatherings or traveling period (range: 0.5 to 5 days), the proportion of the population attending the mass gathering events or on travel (range: 1% to 50%), and the initial reproduction numbers R0 (1.3, 1.5, 1.8) are explored. RESULTS: Mass gatherings that occur within 10 days before the epidemic peak can result in as high as a 10% relative increase in the peak prevalence and the total attack rate, and may have even worse impacts on local communities and travelers' families. Holiday traveling can lead to a second epidemic peak under certain scenarios. Conversely, mass traveling or gatherings may have little effect when occurring much earlier or later than the epidemic peak, e.g., more than 40 days earlier or 20 days later than the peak when the initial R0 = 1.5. CONCLUSIONS: Our results suggest that monitoring, postponing, or cancelling large public gatherings may be warranted close to the epidemic peak but not earlier or later during the epidemic. Influenza activity should also be closely monitored for a potential second peak if holiday traveling occurs when prevalence is high.