Predicting Length of Stay for Obstetric Patients via Electronic Medical Records

Transcription

1 Predicting Length of Stay for Obstetric Patients via Electronic Medical Records Cheng Gao a, Abel N. Kho b, Catherine Ivory c, Sarah Osundson d, Bradley A. Malin a, e, You Chen a a Dept. of Bioedical Inforatics, School of Medicine, Vanderbilt University, Nashville, TN, United States b Institute for Public Health and Medicine, Northwestern University, Chicago, IL, United States c School of Nursing, Vanderbilt University, Nashville, TN, United States d Dept. of Obstetrics and Gynecology, School of Medicine, Vanderbilt University, Nashville, TN, United States e Dept. of Electrical Engineering & Coputer Science, School of Engineering, Vanderbilt University, Nashville, TN, United States Abstract Obstetric care refers to the care provided to patients during ante-, intra-, and postpartu periods. Predicting length of stay (LOS) for these patients during their hospitalizations can assist healthcare organizations to allocate hospital resources ore effectively and efficiently, which, in turn, can iprove aternal care quality and reduce patients costs. In this paper, we investigate the extent to which LOS is associated with a patient s edical history. We introduce a achine learning fraework to incorporate a patient s prior conditions (e.g., diagnostic codes) as features in a predictive odel for LOS. We evaluate the fraework with one-, two- and three-year historical billing data in electronic edical records for 9,88 obstetric patients in a large acadeic edical center. The results show that our fraework achieve an average accuracy of 49.3%, which is higher than the baseline accuracy 37.7% (that relies on a patient s age). Furtherore, the ost predictive features were found to be statistically significant discriinative features. These included historical billing codes for noral delivery (indicative of shorter stay) and antepartu hypertension (indicative of longer stay). Keywords: length of stay (LOS), electronic edical records (EMRs), rando forest, obstetrics, prediction Introduction Electronic edical record (EMR) systes have been widely adopted in the United States (US) [-3] and abroad [,4]. These systes have enabled a substantial aount of patient-specific data to be captured during the routine practice of healthcare organizations (HCOs) [-5]. This inforation is quite heterogeneous, ranging fro structured diagnoses, edication regiens, laboratory test results and vital signs to un- or seistructured clinical narratives. The data stored in EMRs is increasingly being recognized for its ability to support nuerous activities, such as clinical decision aking [5, 9], patient safety iproving [0-] and discovery-driven bioedical research [-4]. One of the ore challenging healthcare anageent environents in odern tie is the safety of aternity. Over the past several decades, the aternal ortality ratio (MMR) has risen draatically in the US. MMR has doubled fro 7. deaths of others per 00,000 live births in 987 to 4 in 05 [6]. At the sae tie, obstetric care is the ost coon and costly type of hospital care for all payers in the US [7-9]. Prediction of the length of stay (LOS) for obstetric patients during their hospitalization can help unit anagers and adinistrators to ake decisions about hospital resource allocation and configuration. This allows for obstetric care iproveent before, during and after childbirth. Better organized care can reduce the orbidity and ortality of woen, as well as newborn babies [8,9], and reduce aternityrelated costs. Moreover, the failies of obstetric patients who are hospitalized frequently inquire about the expected duration of the hospitalization. The incorporation of an accurate estiate of LOS in counseling discussions ay itigate anxieties over the uncertainty of a hospital stay as well as prepare for discharge to hoe or elsewhere [0]. Previous research has focused on explaining factors that lead to LOS variation in general. LOS has, for instance, been shown to be influenced by nuerous factors, including a patient s deographics (e.g., age), socioeconoic status (e.g., incoe, education, and occupation), insurance types (e.g., coercial, private and Medicaid and Medicare) and severity of illnesses [-3]. It has further been shown to be affiliated with HCOspecific factors, such as physicians work efficiency [5,3] and the availability of professional language interpretation services [5,6]. However, the coplex relationships between these factors further exacerbates the coplexity of LOS prediction. Thus, it is challenging to build LOS prediction odels that rely solely on expert knowledge and inforation ascertained at the tie of a patient s adission to a hospital. Thus, this paper presents a pilot study on the feasibility of a patient s historical diagnoses, as docuented in an EMR for LOS predictive odels. This study is predicated on the hypothesis that LOS is related to a patient s edical history. To investigate this hypothesis, we study three years worth of historic diagnosis codes (prior to their ost recent adission) for patients on an obstetric service at Northwestern Meorial Hospital (NMH) in Chicago, Illinois, USA. Specifically, we extracted EMR data in the for of International Classification of Diseases, ninth revision (ICD-9) [7] codes. We designed a achine learning fraework to predict LOS. The results indicate that it predicts LOS within hours with over 0% greater accuracy than baseline odels that rely solely on the patient s deographics at tie of adission. In addition, we show that certain ICD-9 codes were statistically significant in their predictive capability, which suggests they are ripe further investigation and transition into clinical decision support. Methods Figure provides the EMR data and analytics work-flow adopted for this investigation. First, the ICD-9 codes and LOS for patients are extracted fro the EMR. These are subsequently applied to train and test a predictive odel. Finally, the ost discriinant ICD-9 codes are prioritized and assessed for statistical significance.

2 Figure The process by which the LOS predictive odel is coposed and discriinative features are discovered. Dataset The dataset was drawn fro the Cerner inpatient EMR syste in place at NMH fro July 007 to July 0. It includes the following patient-specific features: ) deographics (e.g., age), ) encounter inforation (e.g., adission and discharge date), 3) diagnosis (e.g., billing codes) assigned to an encounter, and 4) clinical (e.g., obstetrics) service to which the patient was assigned. In total, there were 988 inpatients in the dataset with 849 distinct ICD-9 codes. We consider all inpatients on the obstetric service during 00 and 0 for prediction and rely on EMR data between 007 and 00 as features for our odels. The LOS for an encounter was calculated as the hourly difference between adission and discharge. We use a patient s age as a baseline prediction for LOS. Table suaries the average nuber of ICD-9 codes for the investigated patients in one-, two- and three year EMRs, the avarage age of the investigated patients, and the average LOS for these patients on the obstetric service during 00 and 0. Table Statistics of ICD-9 codes, age and LOS # of ICD-9 codes Age LOS year years 3 years Mean Min 4.6 Max The distribution of inpatients on LOS is shown in Figure. The LOS for ost (74%) of obstetric patients are ranging fro 48 to 96 hours. Figure Frequency of LOS for inpatients in the study. Predictive Model We adopted a rando forest odel to predict a patient s LOS according to his historic assigned ICD-9 codes. We rely on a rando forest because it is a useful enseble approach for regression and classification. Specifically, the average LOS fro all the trees is used for prediction. We odel the data as a atrix, as shown in Equation (). Let n be the nuber of patients and be the nuber of unique ICD- 9 codes. In this atrix, each row represents a specific patient and each colun is a specific characteristic of the patient. The first colun is a patient s LOS (continuous variable) and the rest of the coluns are the ICD-9 codes for each patient. To itigate the influence of repeat visits for patients, treat each ICD-9 code as a binary variable, such that it is arked as if the patient was assigned this diagnosis at least once and otherwise 0. C LOS LOS C C C C LOS 3 C 3 C 3 C 3 () LOS n C n C n C n [ LOS n C n C n C n ] To copare perforance, we also define a baseline ethod, which leverages age as a single feature in prediction. Beyond the LOS prediction, the iportance ranking of features in the predictive odel used will be critical for HCOs to apply and interprate the corresponding results. The rando forest regression odel enables iportance ranking for each feature, which is calculated as extent to which prediction error increases when data for the investigated feature is peruted while all others are held constant [8]. Perforance Evaluation C A rando forest regression predicts an LOS as a continuous value. We evaluated the perforance of this prediction with respect to the actual LOS as follows. Let us assue LOS i and LOS i (i =,, n) are the true and predicted LOS value, respectively.we calculate the difference between LOS i and LOS i as t i. If t i is saller than a predefined tolerance threshold τ, we clai a correct prediction for the i th patient:, LOS b i = { i LOS i τ () 0, otherwise The accuracy (Acc) of prediction is thus assessed as n Acc = n (3) Experient Design i= b i This section begins with a description of four coparative odels in our fraework. We, then, describe how paraeters are selected and copare the odels on a range over the paraeters. For the paraeters selection and odels coparisions, we use 5 randoized runs of 3-fold crossvalidation. Finally, we describe a statistical test strategy to deterine the significant influences ofdiscriinant features on the LOS. Coparative Models Model M0 M M M3 Table Coparative Models Description Age One year of ICD-9 codes Two years of ICD-9 codes Three years of ICD-9 codes Table suaries the four odels. The baseline odel M 0 uses age as a lone feature. The other three predictive odels (M to M 3 ) rely on one, two, and three-years worth of historical ICD-9 codes. Paraeter Selection There are three paraeters in the fraework that need to be tuned: ) nuber of trees in the rando forest, ) τ and 3)

3 nuber of ICD-9 codes in the odel. Since the odels M, M and M 3 are the sae in ters of the prediction algorith,,we leverage M 3 as a representative odel to select an optial value for each of the paraeters. Nuber of Trees It has been shown that the perforance of rando forest regression is relatively insensitive to the nuber of trees [8]. However, we need to avoid selecting too sall of a value (that results in poor accuracy) and too large of a value (that leads to excessive coputational load). Thus, we evaluated the nuber of trees for the rando forest. We leverage the distribution of predictive perforance to select an optial nuber. The odel accuracy will grow with the nuber of trees, up to a certain point, which is where we fix the value. LOS Threshold τ The threshold, τ introduced above represents the difference tolerance between predicted LOS and the true LOS and it should be different between HCOs. In this case, the accuracy in LOS prediction is evaluated under a set of thresholds {5,, 4, 36, 48} satisfy the requireents of disparate HCOs. Nuber of Predictors The nuber of ICD-9 codes (i.e., features) in this study is relatively large (i.e., 849 in total). As such, we ai to perfor diensionality reduction and derive ore anageable odel. The specific precedure is as follows. First, we sort all of the predictors on their iportance in descending order. Second, we select a subset of the features and predict LOS. We choose the subset with the sallest size but highest predicting accuracy. Figure 3 The accuracy of odels as a function of the nuber of trees in the rando forest. The selection of τ depends on the requireents of an HCOs. It can be seen in Figure 3 that the accuracy grows with τ. If an HCO can accept the predicted LOS between a range of hours shorter and hours longer than the real LOS, then hours could be selected as the value of τ. To this analysis, we set τ to be hours. Model Evaluation We applied three strategies to evaluate the odels. These are suarized in Table 3. Table 3 Model evaluation strategies Strategy # Trees % Features A Vary Constant Constant B Constant Vary Constant C Constant Constant vary Feature Discriinantion Analysis Two saple t test is used to copare LOS between patients with and without a certain ICD-9 code. We conduct such analysis for the top ten ICD-9 codes with the highest iportances derieved fro the rando forest regression odels. For each investigated ICD-9 code, we will test the significnace of the differences on LOS for patients with and without the code. P value for each significnace test will be provided. Results Model Paraaterization Figure 3 depicts the predictive perforance of the odels on a varying nuber of trees. The perforance was evaluated while varing τ over 5,, 4, 36 and 48 hours. It was observed that the accuracy did not iprove when the nuber of trees increased beyond 50, which is where we fix the nuber of trees in the rando forest. Figure 4 The accuracy of odels as a function of top percent of features (top %, 0%, 0%, 40% and 80%) As shown in Figure 4 odel accuracy reaches its highest level when the odel is based on the top 0% of the features. After this point, accuray reains relatively constant. As such, our odels are based on the top 0% of the features. Model Evaluation Figures 5-7 depict odel perforance as a function of the nuber of trees, LOS threshold and percent of features, respectively. In general, it can be seen that the odels that incorporate ICD-9 codes have better perforance of LOS than the baseline odels. Additionally, odels M, M and M 3 have alost the sae predictive perforance. This iplies that oneyear of historical ICD-9 codes ay be adequate for LOS prediction. The accuracy of the four odels (where τ is set to hours and the feature set is fixed to the top 0%) on a varying nuber of tress (0, 50, 00, 00 and 300) is shown in Figure 5. It can be seen that the accuracy of M, M and M 3 is uch higher than M 0. It can further be seen that the nuber of trees has inial influence on the accuracy of all four odels. In Figure 6, it can be seen that as the LOS threshold increases, the accuracy of all four odels iproves. However, the predictive perforance does not iprove when varying the percent of features (as shown in Figure 7).

4 outperfors M 0 (37.7%). Moreover, for patients with LOS 96 hours, M 3 outperfors M 0 as well. Table 4 Model coparison for all patients and patients with LOS 96 Model All patients Accuracy LOS 96 hours M0 37.7% 0.00% M3 49.3% 5.8% A Figure 5 Accuracy of the odels as a function of the nuber of decision trees with a constant τ ( hours) and set of predictors (top 0%). Feature Discriinantion Analysis Out of 849 ICD-9 codes, 9 were selected for further analysis in ters of their clinical iplication (as shown in Table 5). Each of the codes in the top 0 ranks exhibited a statistically significant influence on the LOS between LOS for patients with and without such codes. Such evidence ay help healthcare organizations (HCOs) to adopt resource allocation strategies that optiize the care anageent and iprove care quality. As an exaple, the ost discriinant ICD-9 code was 650: Noral Delievery. The ean LOS for patients with and without Noral Delievery in their history was 58. and 76.5 hours, respectively. The p-value for this difference was less than As another exaple, patients with ICD-9 code of Elderly ultigravida with antepartu condition or coplication stayed about 0 ore hours than those without that code (8 versus 70.6). B Figure 6 Accuracy of the odels as a function of the LOS threshold with the nuber of trees (50) and set of predictors (top 0%). C Figure 7 Accuracy of the odels as a function of percent of predictors with a constant τ ( hours) and LOS threshold. Finally, to copare the overall accuracy of our odel (M 3 as a representative odel) at their optial paraeters (50 decision trees, hours of LOS threshold, and top 0% features) with the baseline odel, we conducted a series of experients to predict LOS in two settings: ) all patients whose real LOS > 0, which can be leveraged to easure the perforances of odels on patients with varying LOS and ) the subset of patients whose real LOS 96 hours, which can be used for evaluating the perforances of odels on predicting long LOS The results are shown in Table 4. For all investigated patients, M 3 (49.3%) Iportance Rank ICD-9 code Discussion Table 5 Suary for the top 9 ost predictive ICD-9 codes Description The results show that the hisorical inforation in EMRs ay assist in forecasting obstetric patients LOS in a hospital. Specially, our rando forest regression odel predicted LOS with an accuracy of 49% under an error range of hours, which is 0% ore accurate than a baseline ode. The experiental results further deonstrated that a odel based on the top 0% of ICD-9 codes can achieve an accuracy as high as those based on all involved ICD-9 codes (over 800). Additionally, the odels based on the ost recent year of data logged in the EMRs can achieve the siilar perforance with those based on two-, and three-year worth of data. Notably, we also investigated the top 9 ICD-9 codes, which have significant differences in ters of LOS between patients with and without such codes. These results suggest that the HCOs can specialize resource allocation strategies accordingly. Despite the erits of this investigation, we acknowledge that this is a pilot study and there are several liitations. First, the data was collected within a single institution and ay not cover all of a patient s edical history or be readily applicable to another hospital setting. Second, all of the patients in this study were on an obstetric service, such that the fraework ay not be directly extended to other types of patients or healthcare services. Third, the prediction of LOS ay be considered poor (accuracy of 48% in hours tolerance). Other factors that potentially influences LOS can be incorporated in the odel such as certain patient deographics (e.g., race) or physical traits (e.g., height, weight or BMI). Last but not least, if the HCOs can only accept LOS threshold less than hours, the odel built in this study will fail. # of patients (with code verse. without code) Mean LOS (with code verse. without code) p-value 650 Noral delivery 8 vs vs <0.00

5 Elderly ultigravida with antepartu condition or coplication 53 vs vs < V8.8 Encounter for fetal anatoic survey 899 vs vs Unspecified aneia 343 vs vs V8.89 Other specified antenatal screening 580 vs vs V8.4 Antenatal screening for fetal growth retardation using ultrasonics 384 vs vs. 7.3 < V3.9 Unspecified high-risk pregnancy 53 vs vs < V3.8 Supervision of high-risk pregnancy of elderly ultigravida 37 vs vs Unspecified hypertension antepartu 89 vs vs. 7.6 <0.00 Conclusions This paper assessed the feasibility of a achine learning-based fraework for predicting LOS for obstetric patients using historical diagnoses. We showed that one year worth of diagnostic history can predict hospitalization LOS with accuracy higher than that of a siple baseline. We plan to extend the fraework by including additional types of historical inforation (e.g., edications) and leveraging the chronological order of such inforation. References [] Dranove D, Garthwaite C, Li B, Ody C. Investent subsidies and the adoption of electronic edical records in hospitals. Journal of health econoics. 05;44: [] Chen Y, Ghosh J, Bejan C, Gunter C, Kho A, Liebovitz D, Sun J, Denny J, Malin B. Building bridges across electronic health record systes through inferred phenotypic topics. Journal of Bioedical inforatics. 05; 55: PMID: [3] Yan C, Chen Y, Li B, Liebovitz D, Malin B. Learning Clinical Workflows to Identify Subgroups of Heart Failure Patients. Proceedings of the Aerican Medical Inforatics Annual Fall Syposiu. 06; 06; PMID: 8699 [4] Zheng T, Xie W, Xu L, He X, Zhang Y, You M, Yang G, Chen Y. A achine learning-based fraework to identify type diabetes through electronic health records. International Journal of Medical Inforatics. 07 Jan 3;97:0-7. [5] Chen Y, Xie W, Gunter CA, Liebovitz D, Mehrotra S, Zhang H, Malin B. Inferring clinical workflow efficiency via electronic edical record utilization. InAMIA Annual Syposiu Proceedings 05 (Vol. 05, p. 46). Aerican Medical Inforatics Association. [6] WHO, UNICEF, UNFPA, The World Bank, and the United Nations Population Division. Trends in aternal ortality: 990 to 05. Geneva, World Health Organization, 05. [7] Darstadt GL, Bhutta ZA, et.al. Evidence-based, costeffective interventions: how any newborn babies can we save? The Lancet. 005; 356(9463): [8] Schitt SK, Sneed L, Phibbs CS. Costs of Newborn Care in California: A Population-Based Study. Pediatrics. 006; 7(): [9] Kerber KJ, et.al. Continuu of care for aternal, newborn, and child health: fro slogan to service delivery. Lancet. 007; 370(9595): [0] Turner M, Winefield H, Chur-Hansen A. The eotional experiences and supports for parents with babies in a neonatal nursery. Advances in Neonatal Care. 03;3(6): [] Verburg IWM, de Keizer NF, de Jonge E, Peek N. Coparison of regression ethods for odeling intensive care length of stay. PLoS One. 04;9(0). [] Goldfarb MG, Hornbrook MC, Higgins CS. Deterinants of hospital use: a cross-diagnostic analysis. Medical Care. 983: [3] Epstein AM, Stern RS, Weissan JS. Do the poor cost ore? A ultihospital study of patients' socioeconoic status and use of hospital resources. New England Journal of Medicine. 990;3(6):-8. [4] Burns LR, Wholey DR. The effects of patient, hospital, and physician characteristics on length of stay and ortality. Medical care. 99;9(3):5-7. [5] Federan EJ, Drebing CE, Boisvert C, Penk W, Binus G, Rosenheck R. Relationship between cliate and psychiatric inpatient length of stay in Veterans Health Adinistration hospitals. Aerican Journal of Psychiatry. 000;57(0): [6] Lindhol M, Hargraves JL, Ferguson WJ, Reed G. Professional language interpretation and inpatient length of stay and readission rates. Journal of general internal edicine. 0;7(0):94-9. [7] Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical coorbidity index for use with ICD-9-CM adinistrative databases. Journal of clinical epideiology. 99;45(6):63-9. [8] Liaw A, Wiener M. Classification and regression by randoforest. R news. 00;(3):8-. [9] Chen Y, Lorenzi N, Sandberg W, Wolgast K, Malin B. Identifying Collaborative Care Teas through Electronic Medical Record Utilization Patterns. Journal of the Aerican Medical Inforatics Association. 07; 4 (e): e-e0. PMID: [0] Chen Y, Nyeba S, Malin B. Auditing edical record accesses via healthcare interaction networks. Proceedings of the Aerican Medical Inforatics Association Annual Syposiu. 0; [] Chen Y, Nyeba S, Malin B. Detecting anoalous insiders in collaborative inforation systes. IEEE Transactions on Dependable and Secure Coputing. 0; 9(3): Address for correspondence Cheng Gao: cheng.gao@vanderbilt.edu