Samantha Thomas, MB

Hospital discharge planning is a complex process requiring efficient coordination of many different medical and social support services. For this reason, multidisciplinary teams work together to develop individualized discharge plans in an attempt to reduce preventable adverse events related to hospital discharge.[1, 2, 3, 4, 5] Despite these ongoing efforts, optimal discharge strategies have yet to be realized.[1, 4, 5, 6, 7, 8, 9]

One factor that may improve the discharge process is the early identification of patients who are approaching discharge.[10] Multidisciplinary teams cannot fully deploy comprehensive discharge plans until a physician deems a patient to be approaching discharge readiness.[8]

To our knowledge, no studies have examined the performance of physician predictions of upcoming discharge. Instead, prior studies have found that physicians have difficulty predicting the length of stay for patients seen in the emergency room and for elderly patients newly admitted to general medicine floor.[11, 12] The purpose of this study was to evaluate the ability of inpatient general medicine physicians to predict next or same‐day hospital discharges to help inform the timing of discharge planning.

METHODS

We collected daily in‐person predictions from all senior residents and attendings separately on the inpatient general medicine teams (5 resident/attending services and 4 attending‐only services) at a single 950‐bed academic medical center. We asked these physicians to predict whether each patient under their care had a greater than or equal to 80% chance of being discharged the next day, the same day, or neither (ie, no discharge on the next or same day).

Physician predictions of discharge occurred Monday through Friday at 1 of 3 time points: morning (79 _am), midday (122 _pm), or afternoon (57 _pm). Data collection focused on 1 time point per week during 2 different weeks in November 2013 and 1 week in February 2014. Predictions of same‐day discharge could only be made at the morning and midday time points. Each patient could have multiple predictions if they remained hospitalized during subsequent assessments. For each physician making a prediction, we recorded the physician training level (resident or attending).

This protocol was deemed exempt by our university's institutional review board.

RESULTS

A total of 2660 predictions were made by 24 attendings and 15 residents. Nineteen predictions were excluded because of missing prediction type or date of discharge, leaving 2641 predictions for analysis. Table 1 summarizes the total number of predictions within subgroups.

The overall daily discharge rate in our population was 22.3% (see Supporting Table 1 in the online version of this article for the raw values). The SN and PPV of physician predictions of next‐day discharge were 48% (95% confidence interval [CI]: 43%‐52%) and 51% (95% CI: 46%‐56%), respectively. The SN and PPV for same‐day discharge predictions were 73% (95% CI: 68%‐78%) and 69% (95% CI: 64%‐73%), respectively. The SP for next and same‐day discharge predictions was 90% (95% CI: 89%‐91%) and 95% (95% CI: 94%‐96%), whereas the NPV was 89% (95% CI: 88%‐90%) and 96% (95% CI: 95%‐97%), respectively.

Outcome measures for each prediction type are stratified by time of day and summarized in Table 2. For next‐day discharge predictions, the SN and PPV were lowest in the morning (SN 27%, PPV 33%) and peaked by the afternoon (SN 67%, PPV 69%). Similarly, for same‐day discharges, SN and PPV were highest later in the day (midday SN 88%, PPV 79%). This trend is also demonstrated in the SP and NPV, which increased as time to actual discharge approached, although the trends are not as pronounced as for SN and PPV.

The overall agreement between resident and attending predictions was measured and found to have kappa values of 0.51 (P < 0.001) for next‐day predictions and 0.73 (P < 0.001) for same‐day predictions, indicating moderate and substantial agreement, respectively (see Supporting Table 2 in the online version of this article).[13]

DISCUSSION

This is the first study, to our knowledge, to examine the ability of physicians to predict upcoming discharge during the course of routine general medicine inpatient care. We found that although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges showed continual improvement as the difference between the prediction time and time of actual discharge shortened.

For next‐day predictions, the most accurate time point was the afternoon, when physicians correctly predicted more than two‐thirds of actual next‐day discharges. This finding suggests that physicians can provide meaningful discharge estimates as early as the afternoon prior to expected discharge. This may be an optimal time for physicians to meet with the multidisciplinary discharge teams, as many preparations hinge on timely and accurate predictions of discharge (eg, arranging patient transportation, postdischarge visits by a home health company). Multidisciplinary teams will also be reassured that an afternoon prediction of next‐day discharge would only prematurely activate discharge resources in roughly 3 out of every 10 occurrences. Even in these instances, patients may benefit from the extra time for disease counselling, medication teaching, and arrangement of home services.[4, 5, 6, 7, 8, 9]

This investigation has several limitations. Our study was conducted at a large tertiary care center over brief time periods, with an overall discharge rate of about 1 in 5 patients per day. Thus, the results may not be generalizable to hospitals with different patient populations, volume, or turnover, or when predictions are made at different times throughout the year. Furthermore, we were unable to determine if the outcome measures were affected by prolonged lengths of stay or excessive predictions on relatively few patients. However, we sought to mitigate these constraints by surveying many different respondents with varying experience levels, caring for a heterogeneous patient population at nonconsecutive time points during the year. A review of our hospital's administrative data suggests that the bed occupancy and average length of stay during our surveys were similar with most other time points during the year, and therefore representative of a typical inpatient general medicine service.

Our investigation was a novel investigation into the performance of physician discharge predictions, which are daily predictions made either explicitly or implicitly by physicians caring for patients on a general medicine ward. By utilizing a simple, subjective survey without bulky calculations, this approach closely mirrors real‐world practice patterns, and if further validated, could be easily assimilated into the normal workflow of a wide range of busy clinicians to more effectively activate individualized discharge plans.[1, 2, 3, 4, 5]

Future work could capture additional patient information, such as functional status, diagnosis, and current length of stay, which would allow identification of certain subsets of patients in which physicians are more or less accurate in predicting hospital discharge. Additionally, the outcomes of incorrect predictions, particularly the surprise discharges that left even though they were predicted to stay, could be assessed. If patients were discharged prematurely, this may be reflected by a higher 30‐day readmission rate, lower clinic follow‐up rate, and/or lower patient satisfaction scores.

CONCLUSION

Although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges steadily improves as the difference between the prediction time and time of actual discharge shortened. It remains to be determined if systematic incorporation of physician discharge predictions into standard workflows will improve the effectiveness of transition of care interventions.

Disclosure: Nothing to report.

Hospital discharge planning is a complex process requiring efficient coordination of many different medical and social support services. For this reason, multidisciplinary teams work together to develop individualized discharge plans in an attempt to reduce preventable adverse events related to hospital discharge.[1, 2, 3, 4, 5] Despite these ongoing efforts, optimal discharge strategies have yet to be realized.[1, 4, 5, 6, 7, 8, 9]

One factor that may improve the discharge process is the early identification of patients who are approaching discharge.[10] Multidisciplinary teams cannot fully deploy comprehensive discharge plans until a physician deems a patient to be approaching discharge readiness.[8]

To our knowledge, no studies have examined the performance of physician predictions of upcoming discharge. Instead, prior studies have found that physicians have difficulty predicting the length of stay for patients seen in the emergency room and for elderly patients newly admitted to general medicine floor.[11, 12] The purpose of this study was to evaluate the ability of inpatient general medicine physicians to predict next or same‐day hospital discharges to help inform the timing of discharge planning.

METHODS

We collected daily in‐person predictions from all senior residents and attendings separately on the inpatient general medicine teams (5 resident/attending services and 4 attending‐only services) at a single 950‐bed academic medical center. We asked these physicians to predict whether each patient under their care had a greater than or equal to 80% chance of being discharged the next day, the same day, or neither (ie, no discharge on the next or same day).

Physician predictions of discharge occurred Monday through Friday at 1 of 3 time points: morning (79 _am), midday (122 _pm), or afternoon (57 _pm). Data collection focused on 1 time point per week during 2 different weeks in November 2013 and 1 week in February 2014. Predictions of same‐day discharge could only be made at the morning and midday time points. Each patient could have multiple predictions if they remained hospitalized during subsequent assessments. For each physician making a prediction, we recorded the physician training level (resident or attending).

This protocol was deemed exempt by our university's institutional review board.

RESULTS

A total of 2660 predictions were made by 24 attendings and 15 residents. Nineteen predictions were excluded because of missing prediction type or date of discharge, leaving 2641 predictions for analysis. Table 1 summarizes the total number of predictions within subgroups.

The overall daily discharge rate in our population was 22.3% (see Supporting Table 1 in the online version of this article for the raw values). The SN and PPV of physician predictions of next‐day discharge were 48% (95% confidence interval [CI]: 43%‐52%) and 51% (95% CI: 46%‐56%), respectively. The SN and PPV for same‐day discharge predictions were 73% (95% CI: 68%‐78%) and 69% (95% CI: 64%‐73%), respectively. The SP for next and same‐day discharge predictions was 90% (95% CI: 89%‐91%) and 95% (95% CI: 94%‐96%), whereas the NPV was 89% (95% CI: 88%‐90%) and 96% (95% CI: 95%‐97%), respectively.

Outcome measures for each prediction type are stratified by time of day and summarized in Table 2. For next‐day discharge predictions, the SN and PPV were lowest in the morning (SN 27%, PPV 33%) and peaked by the afternoon (SN 67%, PPV 69%). Similarly, for same‐day discharges, SN and PPV were highest later in the day (midday SN 88%, PPV 79%). This trend is also demonstrated in the SP and NPV, which increased as time to actual discharge approached, although the trends are not as pronounced as for SN and PPV.

The overall agreement between resident and attending predictions was measured and found to have kappa values of 0.51 (P < 0.001) for next‐day predictions and 0.73 (P < 0.001) for same‐day predictions, indicating moderate and substantial agreement, respectively (see Supporting Table 2 in the online version of this article).[13]

DISCUSSION

This is the first study, to our knowledge, to examine the ability of physicians to predict upcoming discharge during the course of routine general medicine inpatient care. We found that although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges showed continual improvement as the difference between the prediction time and time of actual discharge shortened.

For next‐day predictions, the most accurate time point was the afternoon, when physicians correctly predicted more than two‐thirds of actual next‐day discharges. This finding suggests that physicians can provide meaningful discharge estimates as early as the afternoon prior to expected discharge. This may be an optimal time for physicians to meet with the multidisciplinary discharge teams, as many preparations hinge on timely and accurate predictions of discharge (eg, arranging patient transportation, postdischarge visits by a home health company). Multidisciplinary teams will also be reassured that an afternoon prediction of next‐day discharge would only prematurely activate discharge resources in roughly 3 out of every 10 occurrences. Even in these instances, patients may benefit from the extra time for disease counselling, medication teaching, and arrangement of home services.[4, 5, 6, 7, 8, 9]

This investigation has several limitations. Our study was conducted at a large tertiary care center over brief time periods, with an overall discharge rate of about 1 in 5 patients per day. Thus, the results may not be generalizable to hospitals with different patient populations, volume, or turnover, or when predictions are made at different times throughout the year. Furthermore, we were unable to determine if the outcome measures were affected by prolonged lengths of stay or excessive predictions on relatively few patients. However, we sought to mitigate these constraints by surveying many different respondents with varying experience levels, caring for a heterogeneous patient population at nonconsecutive time points during the year. A review of our hospital's administrative data suggests that the bed occupancy and average length of stay during our surveys were similar with most other time points during the year, and therefore representative of a typical inpatient general medicine service.

Our investigation was a novel investigation into the performance of physician discharge predictions, which are daily predictions made either explicitly or implicitly by physicians caring for patients on a general medicine ward. By utilizing a simple, subjective survey without bulky calculations, this approach closely mirrors real‐world practice patterns, and if further validated, could be easily assimilated into the normal workflow of a wide range of busy clinicians to more effectively activate individualized discharge plans.[1, 2, 3, 4, 5]

Future work could capture additional patient information, such as functional status, diagnosis, and current length of stay, which would allow identification of certain subsets of patients in which physicians are more or less accurate in predicting hospital discharge. Additionally, the outcomes of incorrect predictions, particularly the surprise discharges that left even though they were predicted to stay, could be assessed. If patients were discharged prematurely, this may be reflected by a higher 30‐day readmission rate, lower clinic follow‐up rate, and/or lower patient satisfaction scores.

CONCLUSION

Although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges steadily improves as the difference between the prediction time and time of actual discharge shortened. It remains to be determined if systematic incorporation of physician discharge predictions into standard workflows will improve the effectiveness of transition of care interventions.

Disclosure: Nothing to report.

Hospital discharge planning is a complex process requiring efficient coordination of many different medical and social support services. For this reason, multidisciplinary teams work together to develop individualized discharge plans in an attempt to reduce preventable adverse events related to hospital discharge.[1, 2, 3, 4, 5] Despite these ongoing efforts, optimal discharge strategies have yet to be realized.[1, 4, 5, 6, 7, 8, 9]

One factor that may improve the discharge process is the early identification of patients who are approaching discharge.[10] Multidisciplinary teams cannot fully deploy comprehensive discharge plans until a physician deems a patient to be approaching discharge readiness.[8]

To our knowledge, no studies have examined the performance of physician predictions of upcoming discharge. Instead, prior studies have found that physicians have difficulty predicting the length of stay for patients seen in the emergency room and for elderly patients newly admitted to general medicine floor.[11, 12] The purpose of this study was to evaluate the ability of inpatient general medicine physicians to predict next or same‐day hospital discharges to help inform the timing of discharge planning.

METHODS

We collected daily in‐person predictions from all senior residents and attendings separately on the inpatient general medicine teams (5 resident/attending services and 4 attending‐only services) at a single 950‐bed academic medical center. We asked these physicians to predict whether each patient under their care had a greater than or equal to 80% chance of being discharged the next day, the same day, or neither (ie, no discharge on the next or same day).

Physician predictions of discharge occurred Monday through Friday at 1 of 3 time points: morning (79 _am), midday (122 _pm), or afternoon (57 _pm). Data collection focused on 1 time point per week during 2 different weeks in November 2013 and 1 week in February 2014. Predictions of same‐day discharge could only be made at the morning and midday time points. Each patient could have multiple predictions if they remained hospitalized during subsequent assessments. For each physician making a prediction, we recorded the physician training level (resident or attending).

This protocol was deemed exempt by our university's institutional review board.

RESULTS

A total of 2660 predictions were made by 24 attendings and 15 residents. Nineteen predictions were excluded because of missing prediction type or date of discharge, leaving 2641 predictions for analysis. Table 1 summarizes the total number of predictions within subgroups.

The overall daily discharge rate in our population was 22.3% (see Supporting Table 1 in the online version of this article for the raw values). The SN and PPV of physician predictions of next‐day discharge were 48% (95% confidence interval [CI]: 43%‐52%) and 51% (95% CI: 46%‐56%), respectively. The SN and PPV for same‐day discharge predictions were 73% (95% CI: 68%‐78%) and 69% (95% CI: 64%‐73%), respectively. The SP for next and same‐day discharge predictions was 90% (95% CI: 89%‐91%) and 95% (95% CI: 94%‐96%), whereas the NPV was 89% (95% CI: 88%‐90%) and 96% (95% CI: 95%‐97%), respectively.

Outcome measures for each prediction type are stratified by time of day and summarized in Table 2. For next‐day discharge predictions, the SN and PPV were lowest in the morning (SN 27%, PPV 33%) and peaked by the afternoon (SN 67%, PPV 69%). Similarly, for same‐day discharges, SN and PPV were highest later in the day (midday SN 88%, PPV 79%). This trend is also demonstrated in the SP and NPV, which increased as time to actual discharge approached, although the trends are not as pronounced as for SN and PPV.

The overall agreement between resident and attending predictions was measured and found to have kappa values of 0.51 (P < 0.001) for next‐day predictions and 0.73 (P < 0.001) for same‐day predictions, indicating moderate and substantial agreement, respectively (see Supporting Table 2 in the online version of this article).[13]

DISCUSSION

This is the first study, to our knowledge, to examine the ability of physicians to predict upcoming discharge during the course of routine general medicine inpatient care. We found that although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges showed continual improvement as the difference between the prediction time and time of actual discharge shortened.

For next‐day predictions, the most accurate time point was the afternoon, when physicians correctly predicted more than two‐thirds of actual next‐day discharges. This finding suggests that physicians can provide meaningful discharge estimates as early as the afternoon prior to expected discharge. This may be an optimal time for physicians to meet with the multidisciplinary discharge teams, as many preparations hinge on timely and accurate predictions of discharge (eg, arranging patient transportation, postdischarge visits by a home health company). Multidisciplinary teams will also be reassured that an afternoon prediction of next‐day discharge would only prematurely activate discharge resources in roughly 3 out of every 10 occurrences. Even in these instances, patients may benefit from the extra time for disease counselling, medication teaching, and arrangement of home services.[4, 5, 6, 7, 8, 9]

This investigation has several limitations. Our study was conducted at a large tertiary care center over brief time periods, with an overall discharge rate of about 1 in 5 patients per day. Thus, the results may not be generalizable to hospitals with different patient populations, volume, or turnover, or when predictions are made at different times throughout the year. Furthermore, we were unable to determine if the outcome measures were affected by prolonged lengths of stay or excessive predictions on relatively few patients. However, we sought to mitigate these constraints by surveying many different respondents with varying experience levels, caring for a heterogeneous patient population at nonconsecutive time points during the year. A review of our hospital's administrative data suggests that the bed occupancy and average length of stay during our surveys were similar with most other time points during the year, and therefore representative of a typical inpatient general medicine service.

Our investigation was a novel investigation into the performance of physician discharge predictions, which are daily predictions made either explicitly or implicitly by physicians caring for patients on a general medicine ward. By utilizing a simple, subjective survey without bulky calculations, this approach closely mirrors real‐world practice patterns, and if further validated, could be easily assimilated into the normal workflow of a wide range of busy clinicians to more effectively activate individualized discharge plans.[1, 2, 3, 4, 5]

Future work could capture additional patient information, such as functional status, diagnosis, and current length of stay, which would allow identification of certain subsets of patients in which physicians are more or less accurate in predicting hospital discharge. Additionally, the outcomes of incorrect predictions, particularly the surprise discharges that left even though they were predicted to stay, could be assessed. If patients were discharged prematurely, this may be reflected by a higher 30‐day readmission rate, lower clinic follow‐up rate, and/or lower patient satisfaction scores.

CONCLUSION

Although physicians are poor predictors of discharge in the morning prior to the day of expected discharge, their ability to correctly predict inpatient discharges steadily improves as the difference between the prediction time and time of actual discharge shortened. It remains to be determined if systematic incorporation of physician discharge predictions into standard workflows will improve the effectiveness of transition of care interventions.

Disclosure: Nothing to report.

High utilizers of hospital services are medically complex, psychosocially vulnerable, and at risk for adverse health outcomes.[1, 2] They make up a fraction of the patient population but use a disproportionate amount of resources, with high rates of emergency department (ED) visits and hospital admissions.[1, 3, 4] Less than 1% of patients account for 21% of national healthcare spending, and hospital costs are the largest category of national healthcare expenditures.[2, 5] Many patients who disproportionately contribute to high healthcare costs also have high hospital admission rates.[6]

Interventions targeting high utilizers have typically focused on the outpatient setting.[7, 8, 9, 10] Interventions using individualized care plans in the ED reduced ED visits from 33% to 70%, but all have required an additional case management program or partnership with an outside nonprofit case management organization.[11, 12, 13] One study by a hospitalist group using individualized care plans reduced ED visits and admissions by 70%, 2 months after care‐plan implementation; however, all of their care plans were focused explicitly on restricting intravenous opiate use for patients with chronic pain.[14]

Given the current focus on cost‐conscious, high‐quality care in the American healthcare system, we designed a quality‐improvement (QI) intervention using individualized care plans to reduce unnecessary healthcare service utilization and hospital costs for the highest utilizers of ED and inpatient care. Our approach focuses on integrating care plans within our electronic medical record (EMR) and implementing them using the existing healthcare workforce. We analyzed pre‐ and postintervention data to determine its effect on service utilization and hospital costs across a regional health system.

METHODS

RESULTS

Table 1 shows the demographics and comorbidities for the 24 patients with care plans included in this analysis. The average age of patients is 38.5 years (range, 2565 years) and a nearly even split between males (11) and females (13). Chronic disease burden is high. Furthermore, 83% of patients have chronic pain and 96% have mental health problems or substance abuse.

Table 2 shows inpatient and ED utilization patterns before and after care‐plan implementation. Inpatient admissions decreased by 56% for the 6 months after care‐plan implementation (P<0.001) and by 50.5% for the 12 months after care‐plan implementation (P=0.003). This translates to a decrease in the average number of admissions per patient from 6.3 to 2.4, 6 months postcare plan, and from 10.9 to 4.8, 12 months postcare plan.

Thirty‐day readmissions also significantly decreased after care‐plan implementation. Among the 23 patients with data 6 months pre and postcare plan, there were 130 readmissions before and 44 readmissions after care‐plan implementation, a 66% reduction (P<0.001). Among the 12 patients with data 12 months pre and postcare plan, there were 106 readmissions before and 45 readmissions after care‐plan implementation, a 51.5% reduction (P=0.002). Inpatient LOS did not show a statistically significant change after care‐plan implementation.

ED visits were similar for the 6 months pre compared to 6 months postcare plan. ED visits at 12 months postcare plan increased from an average of 15.4 visits pre to 25.6 visits per patient postcare plan. This was driven by a single homeless patient with dialysis‐dependent end‐stage renal disease, who had 134 ED visits in the 12 months after careplan implementation. Analysis of the data with this outlier removed showed a reduction in ED visits from an average of 12.3 visits per patient to 10.1 visits per patient in the 12 months postcare plan; however, this was not statistically significant (P=0.66, data not shown).

Table 3 shows inpatient and ED VDCs before and after care‐plan implementation. The average VDCs per patient per admission decreased from $29,852.71 to $15,587.84, 6 months after care‐plan implementation, a 47.7% reduction (P=0.001). The average VDCs per patient per admission decreased from $44,881.66 to $24,958.42, 12 months after care‐plan implementation, a 35.8% reduction (P=0.052). ED costs did not show a statistically significant decrease. However, with the outlier removed as above, costs did decrease by 12.3%, 6 months after care‐plan implementation, approaching statistical significance (P=0.073, data not shown). Combined inpatient and ED variable direct costs decreased by an average of $15,117.30, 6 months after care‐plan implementation, a 45.3% reduction (P=0.002), and by an average of $18,604.50, 12 months after care‐plan implementation, a 25.5% reduction, although this did not reach statistical significance (P=0.129).

DISCUSSION

A multidisciplinary team at our academic medical center developed individualized care plans tailored to the specific medical and psychosocial complexities of high utilizers to reduce unnecessary service utilization and hospital costs. Postintervention analysis shows reduced inpatient admissions and 30‐day readmissions among this population by 50%. Furthermore, inpatient variable direct costs decreased by 47% for the 6 months following care‐plan implementation and by 35% for the 12 months following care‐plan implementation. This translates into a $347,696.97 cost savings for the 23 patients 6 months after care‐plan implementation, and a $223,253.89 cost savings for the 12 patients 12 months after care‐plan implementation. This reduction in utilization and cost was seen across all 3 hospitals in the Duke University Health System, including the only 2 hospitals in Durham, North Carolina. Unlike other urban areas, public transportation in our region is scarce, and the options for hospital shopping in central North Carolina are relatively limited. Although this study does not measure utilization in surrounding counties, we do not feel this occurred as we did not see a rise in requests for medical records nor attempts to contact Duke providers for questions on these patients as a result of our intervention. This, along with our regional health system outcome analysis, provides support that our intervention did not cause patients to seek care elsewhere and result in cost‐shifting to other facilities.

We hypothesize that our care plans may be responsible for decreased admissions and 30‐day readmissions through several mechanisms. By raising awareness of these patients' excessive hospital utilization patterns and making this information readily available through our EMR, providers in the ED may be more conscientious about their admission decisions. Problems that at face value seem acute, are often more chronic and can be better managed in the outpatient setting. Several care plans also explicitly recommend limiting unnecessary intravenous opiate use for chronic pain patients. Other patients who have frequent admissions actually have end‐stage disease, and care plans for these patients help facilitate referrals to hospice programs.

Care plans provide a consistent message of patient histories, utilization patterns, and management strategies, and also serve as a communication tool between hospitalists and ED providers. A systematic review of all ED‐based interventions for high utilizers revealed that most studies did show a reduction in ED visits, but all incorporated case management programs to do so.[15] We did not reduce ED visits, possibly because we lacked the resources and care coordination a community‐based case management program provides. However, care plans did serve as a platform with which hospitalists and ED providers can help coordinate care among multiple outpatient providers. This has potentially limited admissions by providing a referral destination or outpatient point of contact for ED providers. For example, as a result of our intervention, referral mechanisms to our comprehensive pain clinic and outpatient psychiatry clinic have both been strengthened and streamlined. The fact that care plans decreased admissions and readmissions, but not ED visits, suggests that our intervention may not have actually changed patient behavior, but instead changed provider practices in relation to disposition decisions in the ED.

Our QI intervention has several strengths. First, it is fully integrated within our existing healthcare workforce, without the need for an extra case management system. Second, it is seamlessly incorporated into our EMR and represents another potential use of an EMR that has not been previously touted. Third, the multidisciplinary nature of the CCPC ensures that all stakeholders involved in the care of high utilizers are represented. Fourth, the outcome analysis across all 3 hospitals in our health system provides a balancing metric against the notion that our intervention simply caused patients to seek care elsewhere in the region. Last, the QI design and lack of strict inclusion and exclusion criteria adds practicality and shows effectiveness, not just efficacy, of the intervention.

Because this was developed as a QI intervention without strict inclusion and exclusion criteria, generalizability is lacking. In the future, one could use the EMR to more systematically identify high‐utilizing, complex patients. One study showed the ability to use the EMR with a standardized framework to identify hot spotting (high utilizers) and contextual anomaly detection (ie, anomalous utilization cases where patient‐incurred levels of utilization are unexpected given their clinical characteristics).[16] The nonrandomized, retrospective pre/post‐intervention analysis without a control group diminishes the external validity of the results and does introduce the potential for bias.

One of the primary study limitations includes the small sample size of only 24 patients. Admittedly, these first 24 patients are the absolute highest utilizers of care at our hospital, possibly making their utilization patterns more amenable to our intervention. The 96% prevalence rate of mental health and substance abuse in our cohort is significantly higher than other published data among high utilizers.[4, 17, 18] We are continuing to develop care plans for additional high‐utilizing, complex patients, and expect to enroll more patients with end‐stage disease, and relatively fewer with substance abuse or psychiatric illness as time goes on. It is possible this new cohort of patients has proportionally less unnecessary utilization, thus limiting our intervention effect. One final limitation of our study is the lack of care quality and patient safety outcomes. In future studies, health outcomes, adverse events, and outpatient care utilization will be important balancing measures to include.

In conclusion, we showed that a QI intervention using individualized care plans reduces hospital admissions, 30‐day readmissions, and hospital costs across a regional health system for a group of complex, high‐utilizing patients. This intervention can, and should, be developed by a multidisciplinary team and fully integrated into the existing healthcare workforce and EMR to ensure appropriateness, effectiveness, and longevity. Going forward, it will be imperative to evaluate this intervention prospectively, at multiple sites, in coordination with outpatient providers, and including quality and safety outcomes to determine if this hospital‐based intervention impacts care coordination, utilization rates, cost, and health outcomes across the broader healthcare system.

High utilizers of hospital services are medically complex, psychosocially vulnerable, and at risk for adverse health outcomes.[1, 2] They make up a fraction of the patient population but use a disproportionate amount of resources, with high rates of emergency department (ED) visits and hospital admissions.[1, 3, 4] Less than 1% of patients account for 21% of national healthcare spending, and hospital costs are the largest category of national healthcare expenditures.[2, 5] Many patients who disproportionately contribute to high healthcare costs also have high hospital admission rates.[6]

Interventions targeting high utilizers have typically focused on the outpatient setting.[7, 8, 9, 10] Interventions using individualized care plans in the ED reduced ED visits from 33% to 70%, but all have required an additional case management program or partnership with an outside nonprofit case management organization.[11, 12, 13] One study by a hospitalist group using individualized care plans reduced ED visits and admissions by 70%, 2 months after care‐plan implementation; however, all of their care plans were focused explicitly on restricting intravenous opiate use for patients with chronic pain.[14]

Given the current focus on cost‐conscious, high‐quality care in the American healthcare system, we designed a quality‐improvement (QI) intervention using individualized care plans to reduce unnecessary healthcare service utilization and hospital costs for the highest utilizers of ED and inpatient care. Our approach focuses on integrating care plans within our electronic medical record (EMR) and implementing them using the existing healthcare workforce. We analyzed pre‐ and postintervention data to determine its effect on service utilization and hospital costs across a regional health system.

METHODS

RESULTS

Table 1 shows the demographics and comorbidities for the 24 patients with care plans included in this analysis. The average age of patients is 38.5 years (range, 2565 years) and a nearly even split between males (11) and females (13). Chronic disease burden is high. Furthermore, 83% of patients have chronic pain and 96% have mental health problems or substance abuse.

Table 2 shows inpatient and ED utilization patterns before and after care‐plan implementation. Inpatient admissions decreased by 56% for the 6 months after care‐plan implementation (P<0.001) and by 50.5% for the 12 months after care‐plan implementation (P=0.003). This translates to a decrease in the average number of admissions per patient from 6.3 to 2.4, 6 months postcare plan, and from 10.9 to 4.8, 12 months postcare plan.

Thirty‐day readmissions also significantly decreased after care‐plan implementation. Among the 23 patients with data 6 months pre and postcare plan, there were 130 readmissions before and 44 readmissions after care‐plan implementation, a 66% reduction (P<0.001). Among the 12 patients with data 12 months pre and postcare plan, there were 106 readmissions before and 45 readmissions after care‐plan implementation, a 51.5% reduction (P=0.002). Inpatient LOS did not show a statistically significant change after care‐plan implementation.

ED visits were similar for the 6 months pre compared to 6 months postcare plan. ED visits at 12 months postcare plan increased from an average of 15.4 visits pre to 25.6 visits per patient postcare plan. This was driven by a single homeless patient with dialysis‐dependent end‐stage renal disease, who had 134 ED visits in the 12 months after careplan implementation. Analysis of the data with this outlier removed showed a reduction in ED visits from an average of 12.3 visits per patient to 10.1 visits per patient in the 12 months postcare plan; however, this was not statistically significant (P=0.66, data not shown).

Table 3 shows inpatient and ED VDCs before and after care‐plan implementation. The average VDCs per patient per admission decreased from $29,852.71 to $15,587.84, 6 months after care‐plan implementation, a 47.7% reduction (P=0.001). The average VDCs per patient per admission decreased from $44,881.66 to $24,958.42, 12 months after care‐plan implementation, a 35.8% reduction (P=0.052). ED costs did not show a statistically significant decrease. However, with the outlier removed as above, costs did decrease by 12.3%, 6 months after care‐plan implementation, approaching statistical significance (P=0.073, data not shown). Combined inpatient and ED variable direct costs decreased by an average of $15,117.30, 6 months after care‐plan implementation, a 45.3% reduction (P=0.002), and by an average of $18,604.50, 12 months after care‐plan implementation, a 25.5% reduction, although this did not reach statistical significance (P=0.129).

DISCUSSION

A multidisciplinary team at our academic medical center developed individualized care plans tailored to the specific medical and psychosocial complexities of high utilizers to reduce unnecessary service utilization and hospital costs. Postintervention analysis shows reduced inpatient admissions and 30‐day readmissions among this population by 50%. Furthermore, inpatient variable direct costs decreased by 47% for the 6 months following care‐plan implementation and by 35% for the 12 months following care‐plan implementation. This translates into a $347,696.97 cost savings for the 23 patients 6 months after care‐plan implementation, and a $223,253.89 cost savings for the 12 patients 12 months after care‐plan implementation. This reduction in utilization and cost was seen across all 3 hospitals in the Duke University Health System, including the only 2 hospitals in Durham, North Carolina. Unlike other urban areas, public transportation in our region is scarce, and the options for hospital shopping in central North Carolina are relatively limited. Although this study does not measure utilization in surrounding counties, we do not feel this occurred as we did not see a rise in requests for medical records nor attempts to contact Duke providers for questions on these patients as a result of our intervention. This, along with our regional health system outcome analysis, provides support that our intervention did not cause patients to seek care elsewhere and result in cost‐shifting to other facilities.

We hypothesize that our care plans may be responsible for decreased admissions and 30‐day readmissions through several mechanisms. By raising awareness of these patients' excessive hospital utilization patterns and making this information readily available through our EMR, providers in the ED may be more conscientious about their admission decisions. Problems that at face value seem acute, are often more chronic and can be better managed in the outpatient setting. Several care plans also explicitly recommend limiting unnecessary intravenous opiate use for chronic pain patients. Other patients who have frequent admissions actually have end‐stage disease, and care plans for these patients help facilitate referrals to hospice programs.

Care plans provide a consistent message of patient histories, utilization patterns, and management strategies, and also serve as a communication tool between hospitalists and ED providers. A systematic review of all ED‐based interventions for high utilizers revealed that most studies did show a reduction in ED visits, but all incorporated case management programs to do so.[15] We did not reduce ED visits, possibly because we lacked the resources and care coordination a community‐based case management program provides. However, care plans did serve as a platform with which hospitalists and ED providers can help coordinate care among multiple outpatient providers. This has potentially limited admissions by providing a referral destination or outpatient point of contact for ED providers. For example, as a result of our intervention, referral mechanisms to our comprehensive pain clinic and outpatient psychiatry clinic have both been strengthened and streamlined. The fact that care plans decreased admissions and readmissions, but not ED visits, suggests that our intervention may not have actually changed patient behavior, but instead changed provider practices in relation to disposition decisions in the ED.

Our QI intervention has several strengths. First, it is fully integrated within our existing healthcare workforce, without the need for an extra case management system. Second, it is seamlessly incorporated into our EMR and represents another potential use of an EMR that has not been previously touted. Third, the multidisciplinary nature of the CCPC ensures that all stakeholders involved in the care of high utilizers are represented. Fourth, the outcome analysis across all 3 hospitals in our health system provides a balancing metric against the notion that our intervention simply caused patients to seek care elsewhere in the region. Last, the QI design and lack of strict inclusion and exclusion criteria adds practicality and shows effectiveness, not just efficacy, of the intervention.

Because this was developed as a QI intervention without strict inclusion and exclusion criteria, generalizability is lacking. In the future, one could use the EMR to more systematically identify high‐utilizing, complex patients. One study showed the ability to use the EMR with a standardized framework to identify hot spotting (high utilizers) and contextual anomaly detection (ie, anomalous utilization cases where patient‐incurred levels of utilization are unexpected given their clinical characteristics).[16] The nonrandomized, retrospective pre/post‐intervention analysis without a control group diminishes the external validity of the results and does introduce the potential for bias.

One of the primary study limitations includes the small sample size of only 24 patients. Admittedly, these first 24 patients are the absolute highest utilizers of care at our hospital, possibly making their utilization patterns more amenable to our intervention. The 96% prevalence rate of mental health and substance abuse in our cohort is significantly higher than other published data among high utilizers.[4, 17, 18] We are continuing to develop care plans for additional high‐utilizing, complex patients, and expect to enroll more patients with end‐stage disease, and relatively fewer with substance abuse or psychiatric illness as time goes on. It is possible this new cohort of patients has proportionally less unnecessary utilization, thus limiting our intervention effect. One final limitation of our study is the lack of care quality and patient safety outcomes. In future studies, health outcomes, adverse events, and outpatient care utilization will be important balancing measures to include.

In conclusion, we showed that a QI intervention using individualized care plans reduces hospital admissions, 30‐day readmissions, and hospital costs across a regional health system for a group of complex, high‐utilizing patients. This intervention can, and should, be developed by a multidisciplinary team and fully integrated into the existing healthcare workforce and EMR to ensure appropriateness, effectiveness, and longevity. Going forward, it will be imperative to evaluate this intervention prospectively, at multiple sites, in coordination with outpatient providers, and including quality and safety outcomes to determine if this hospital‐based intervention impacts care coordination, utilization rates, cost, and health outcomes across the broader healthcare system.

High utilizers of hospital services are medically complex, psychosocially vulnerable, and at risk for adverse health outcomes.[1, 2] They make up a fraction of the patient population but use a disproportionate amount of resources, with high rates of emergency department (ED) visits and hospital admissions.[1, 3, 4] Less than 1% of patients account for 21% of national healthcare spending, and hospital costs are the largest category of national healthcare expenditures.[2, 5] Many patients who disproportionately contribute to high healthcare costs also have high hospital admission rates.[6]

Interventions targeting high utilizers have typically focused on the outpatient setting.[7, 8, 9, 10] Interventions using individualized care plans in the ED reduced ED visits from 33% to 70%, but all have required an additional case management program or partnership with an outside nonprofit case management organization.[11, 12, 13] One study by a hospitalist group using individualized care plans reduced ED visits and admissions by 70%, 2 months after care‐plan implementation; however, all of their care plans were focused explicitly on restricting intravenous opiate use for patients with chronic pain.[14]

Given the current focus on cost‐conscious, high‐quality care in the American healthcare system, we designed a quality‐improvement (QI) intervention using individualized care plans to reduce unnecessary healthcare service utilization and hospital costs for the highest utilizers of ED and inpatient care. Our approach focuses on integrating care plans within our electronic medical record (EMR) and implementing them using the existing healthcare workforce. We analyzed pre‐ and postintervention data to determine its effect on service utilization and hospital costs across a regional health system.

METHODS

RESULTS

Table 1 shows the demographics and comorbidities for the 24 patients with care plans included in this analysis. The average age of patients is 38.5 years (range, 2565 years) and a nearly even split between males (11) and females (13). Chronic disease burden is high. Furthermore, 83% of patients have chronic pain and 96% have mental health problems or substance abuse.

Table 2 shows inpatient and ED utilization patterns before and after care‐plan implementation. Inpatient admissions decreased by 56% for the 6 months after care‐plan implementation (P<0.001) and by 50.5% for the 12 months after care‐plan implementation (P=0.003). This translates to a decrease in the average number of admissions per patient from 6.3 to 2.4, 6 months postcare plan, and from 10.9 to 4.8, 12 months postcare plan.

Thirty‐day readmissions also significantly decreased after care‐plan implementation. Among the 23 patients with data 6 months pre and postcare plan, there were 130 readmissions before and 44 readmissions after care‐plan implementation, a 66% reduction (P<0.001). Among the 12 patients with data 12 months pre and postcare plan, there were 106 readmissions before and 45 readmissions after care‐plan implementation, a 51.5% reduction (P=0.002). Inpatient LOS did not show a statistically significant change after care‐plan implementation.

ED visits were similar for the 6 months pre compared to 6 months postcare plan. ED visits at 12 months postcare plan increased from an average of 15.4 visits pre to 25.6 visits per patient postcare plan. This was driven by a single homeless patient with dialysis‐dependent end‐stage renal disease, who had 134 ED visits in the 12 months after careplan implementation. Analysis of the data with this outlier removed showed a reduction in ED visits from an average of 12.3 visits per patient to 10.1 visits per patient in the 12 months postcare plan; however, this was not statistically significant (P=0.66, data not shown).

Table 3 shows inpatient and ED VDCs before and after care‐plan implementation. The average VDCs per patient per admission decreased from $29,852.71 to $15,587.84, 6 months after care‐plan implementation, a 47.7% reduction (P=0.001). The average VDCs per patient per admission decreased from $44,881.66 to $24,958.42, 12 months after care‐plan implementation, a 35.8% reduction (P=0.052). ED costs did not show a statistically significant decrease. However, with the outlier removed as above, costs did decrease by 12.3%, 6 months after care‐plan implementation, approaching statistical significance (P=0.073, data not shown). Combined inpatient and ED variable direct costs decreased by an average of $15,117.30, 6 months after care‐plan implementation, a 45.3% reduction (P=0.002), and by an average of $18,604.50, 12 months after care‐plan implementation, a 25.5% reduction, although this did not reach statistical significance (P=0.129).

DISCUSSION

A multidisciplinary team at our academic medical center developed individualized care plans tailored to the specific medical and psychosocial complexities of high utilizers to reduce unnecessary service utilization and hospital costs. Postintervention analysis shows reduced inpatient admissions and 30‐day readmissions among this population by 50%. Furthermore, inpatient variable direct costs decreased by 47% for the 6 months following care‐plan implementation and by 35% for the 12 months following care‐plan implementation. This translates into a $347,696.97 cost savings for the 23 patients 6 months after care‐plan implementation, and a $223,253.89 cost savings for the 12 patients 12 months after care‐plan implementation. This reduction in utilization and cost was seen across all 3 hospitals in the Duke University Health System, including the only 2 hospitals in Durham, North Carolina. Unlike other urban areas, public transportation in our region is scarce, and the options for hospital shopping in central North Carolina are relatively limited. Although this study does not measure utilization in surrounding counties, we do not feel this occurred as we did not see a rise in requests for medical records nor attempts to contact Duke providers for questions on these patients as a result of our intervention. This, along with our regional health system outcome analysis, provides support that our intervention did not cause patients to seek care elsewhere and result in cost‐shifting to other facilities.

We hypothesize that our care plans may be responsible for decreased admissions and 30‐day readmissions through several mechanisms. By raising awareness of these patients' excessive hospital utilization patterns and making this information readily available through our EMR, providers in the ED may be more conscientious about their admission decisions. Problems that at face value seem acute, are often more chronic and can be better managed in the outpatient setting. Several care plans also explicitly recommend limiting unnecessary intravenous opiate use for chronic pain patients. Other patients who have frequent admissions actually have end‐stage disease, and care plans for these patients help facilitate referrals to hospice programs.

Care plans provide a consistent message of patient histories, utilization patterns, and management strategies, and also serve as a communication tool between hospitalists and ED providers. A systematic review of all ED‐based interventions for high utilizers revealed that most studies did show a reduction in ED visits, but all incorporated case management programs to do so.[15] We did not reduce ED visits, possibly because we lacked the resources and care coordination a community‐based case management program provides. However, care plans did serve as a platform with which hospitalists and ED providers can help coordinate care among multiple outpatient providers. This has potentially limited admissions by providing a referral destination or outpatient point of contact for ED providers. For example, as a result of our intervention, referral mechanisms to our comprehensive pain clinic and outpatient psychiatry clinic have both been strengthened and streamlined. The fact that care plans decreased admissions and readmissions, but not ED visits, suggests that our intervention may not have actually changed patient behavior, but instead changed provider practices in relation to disposition decisions in the ED.

Our QI intervention has several strengths. First, it is fully integrated within our existing healthcare workforce, without the need for an extra case management system. Second, it is seamlessly incorporated into our EMR and represents another potential use of an EMR that has not been previously touted. Third, the multidisciplinary nature of the CCPC ensures that all stakeholders involved in the care of high utilizers are represented. Fourth, the outcome analysis across all 3 hospitals in our health system provides a balancing metric against the notion that our intervention simply caused patients to seek care elsewhere in the region. Last, the QI design and lack of strict inclusion and exclusion criteria adds practicality and shows effectiveness, not just efficacy, of the intervention.

Because this was developed as a QI intervention without strict inclusion and exclusion criteria, generalizability is lacking. In the future, one could use the EMR to more systematically identify high‐utilizing, complex patients. One study showed the ability to use the EMR with a standardized framework to identify hot spotting (high utilizers) and contextual anomaly detection (ie, anomalous utilization cases where patient‐incurred levels of utilization are unexpected given their clinical characteristics).[16] The nonrandomized, retrospective pre/post‐intervention analysis without a control group diminishes the external validity of the results and does introduce the potential for bias.

One of the primary study limitations includes the small sample size of only 24 patients. Admittedly, these first 24 patients are the absolute highest utilizers of care at our hospital, possibly making their utilization patterns more amenable to our intervention. The 96% prevalence rate of mental health and substance abuse in our cohort is significantly higher than other published data among high utilizers.[4, 17, 18] We are continuing to develop care plans for additional high‐utilizing, complex patients, and expect to enroll more patients with end‐stage disease, and relatively fewer with substance abuse or psychiatric illness as time goes on. It is possible this new cohort of patients has proportionally less unnecessary utilization, thus limiting our intervention effect. One final limitation of our study is the lack of care quality and patient safety outcomes. In future studies, health outcomes, adverse events, and outpatient care utilization will be important balancing measures to include.

In conclusion, we showed that a QI intervention using individualized care plans reduces hospital admissions, 30‐day readmissions, and hospital costs across a regional health system for a group of complex, high‐utilizing patients. This intervention can, and should, be developed by a multidisciplinary team and fully integrated into the existing healthcare workforce and EMR to ensure appropriateness, effectiveness, and longevity. Going forward, it will be imperative to evaluate this intervention prospectively, at multiple sites, in coordination with outpatient providers, and including quality and safety outcomes to determine if this hospital‐based intervention impacts care coordination, utilization rates, cost, and health outcomes across the broader healthcare system.