第一部分 国外老年人预后模型对北京协和医院老年科患者出院1年内全因死亡风险的预测效能

【背景】 全因死亡预测模型有助医生判断老年患者的不良预后，进行医患共同决策，避免低价值医疗，但是，迄今死亡预测模型在中国大多数老年科患者中没有应用。

【目的】 验证并比较国际上用于老年患者的4个预测模型Walter 指数、老年医学预后指数（GPI）、Charlson共病指数（CCI）及FRAIL量表对北京协和医院老年科患者出院1年内全因死亡的预测效能。

【方法】 查询His，连续纳入2016年1月至2021年12月北京协和医院老年科出院患者（≥70岁，有老年综合评估）；排除住院时间≤24h，>60d的患者。记录基线资料：人口统计学、疾病、血化验数值和老年综合评估。2022年8月至2023年2月对出院1年后的患者/家属电话随访，记录患者死亡时间和原因。4个模型的预测效能通过Delong检验比较受试者工作特征曲线下面积（AUC），绘制校准曲线、决策曲线综合评估。

【结果】 共纳入出院患者832例，中位年龄77（74-82）岁，女性54.3%。出院1年内全因死亡率12%。在校正混杂因素后，多元Logistic回归结果显示，随着4个模型分数的增加，出院1年内死亡风险逐渐增加；4个模型的AUC分别为Walter 指数 0.89，CCI 0.80，GPI 0.75，FRAIL Scale 0.75，表示4个预测模型的区分度均较好，Walter指数区分度最高，显著优于其他模型（Delong检验P均<0.05）；校准曲线显示，除了CCI的校准度欠佳，其他3个模型校准度均较好；决策曲线分析4个模型都具有临床实用性，Walter 指数最佳。

【结论】 4个预后模型Walter 指数、老年医学预后指数、Charlson共病指数及FRAIL量表均适用于预测北京协和医院老年科患者出院1年内全因死亡风险。其中疾病联合功能的模型优于共病模型，Walter指数表现最佳。

第二部分 老年科患者出院1年内全因死亡预测模型的开发及验证

【背景】 经我们验证并推荐Walter指数用于中国老年综合科患者出院1年内死亡预测模型，但该模型并未纳入衰弱、内在能力相关变量。

【目的】 开发我国老年科患者出院1年内全因死亡预测模型。

【方法】 选取2016年1月至2021年12月北京协和医院（Peking Union Medical College Hospital，PUMCH）老年科连续出院患者（≥70岁，有老年综合评估）。住院时间≤24h或>60d的患者被排除。以7:3里比例随机分为训练集和内部验证集；选取2019年1月至2021年12月另外3家三甲医院老年综合科连续出院患者作为外部验证集。经统一培训后，收录资料并于2022年8月至2023年2月电话随访。采用最小绝对收缩和选择算法筛选变量，构建Logistic回归模型，绘制列线图，划分死亡风险分层，计算受试者工作特征曲线下面积（AUC），绘制校准曲线、决策曲线，综合评估模型预测效能。

【结果】 共纳入PUMCH老年科患者训练集582人，中位年龄78岁，女性44.5%；内部验证集250人，中位年龄77岁，女性52.8%；外部验证集231人，中位年龄82岁，女性47.2%。3组患者出院1年内死亡率分别为12.5%、10.8%及6.1%。最终开发2个模型，变量均包括衰弱、血红蛋白和白蛋白。模型1包括Charlson共病指数（PUMCH- Geriatric prognostic index，PUMCH-GPI），模型2包括转移性肿瘤（PUMCH-GPI-tumor，PUMCH-GPIt）。2个模型均有较好的区分度，在训练集中PUMCH-GPI AUC=0.88，敏感度83.3%，特异度80.5%；PUMCH-GPIt AUC=0.86，敏感度76.4%，特异度83.9%。校准曲线显示PUMCH-GPIt在3个数据集中校准曲线均贴合理想校准曲线，PUMCH-GPI在训练集和内部验证集的预测死亡率与实际死亡率相当，但在外部验证集中的校准曲线偏离理想校准曲线。

【结论】 我们开发了2个我国老年科患者出院1年内全因死亡预测模型（PUMCH-GPI和PUMCH-GPIt），变量简单易得，区分度较好，有助于制定临床决策。临床推广价值尚需更多验证。

第三部分 比较新模型与国外模型对我国老年科患者出院1年内全因死亡风险的预测效能

【背景】迄今，国内尚无预测老年科患者全因死亡的预测模型，我们基于老年科患者开发了2个新模型：北京协和医院老年医学预后指数（Peking Union Medical College Hospital Geriatric Prognostic Index，PUMCH-GPI）和基于老年科转移性肿瘤患者的PUMCH-GPIt（PUMCH-GPI-tumor），它们的预测效能需要与国际上预测预后模型比较。

【目的】在2个队列中，分别将我们开发的预测模型PUMCH-GPI、PUMCH-GPIt与Walter指数、老年医学预后指数（GPI）、Charlson共病指数（CCI）及FRAIL量表对我国老年科住院患者出院1年内全因死亡的预测效能进行比较。

【方法】选取2016年1月至2021年12月PUMCH老年科连续出院患者（队列A）以及2019年1月至2021年12月3家三甲医院老年综合科连续出院患者（队列B），（≥70岁，有老年综合评估），住院时间≤24h或>60d的患者被排除。医生经统一培训后，收录资料并电话随访出院1年内生存预后。模型的预测效能通过Delong检验比较区分度（AUC）、校准曲线比较校准度、决策曲线分析评价临床实用性进行综合评估。

【结果】队列A纳入832人，中位年龄78岁，女性44.5%；队列B纳入231人，中位年龄82岁，女性47.2%。出院1年死亡率分别为12.0% 及6.1%。区分度方面，PUMCH-GPI AUC 0.88（队列A）、0.86（队列B）；PUMCH-GPIt AUC 0.86（队列A）、0.83（队列B）；不亚于Walter Index AUC 0.89（队列A）、0.91（队列B），在队列A中AUC均显著优于GPI（0.75）、CCI（0.80）、FRAIL量表（0.75）（Delong 检验P均<0.05）。PUMCH-GPIt在2个队列中校准曲线均贴近理想校准曲线，校准度最佳，具有临床实用性，整体效能优于其他模型。PUMCH-GPI在队列A中的校准度优于Walter指数以及其他3个模型，具有临床实用性，但在队列B中的校准度欠佳。

【结论】在老年科患者出院1年内死亡预测模型中，PUMCH-GPI和PUMCH-GPIt预测效能好，不亚于Walter指数，变量简便易得。优于GPI、共病模型CCI和衰弱模型FRAIL量表。可能具有潜在应用价值。外验证人数较少，尚需要在更多医院老年科患者中验证。

PartⅠ  Efficacy of foreign prognostic models for the older adults in predicting all-cause mortality within 1 year of discharge from the geriatric ward at Peking Union Medical College Hospital

【Background】 All-cause mortality prediction models are useful for physicians to predict poor prognosis of older patients. However, these models have not been applied to most older inpatients in the department of geriatrics in China.

【Objective】 To validate and compare the performance of 4 prediction models, Walter Index, Geriatric Prognostic Index (GPI), Charlson Comorbidity Index (CCI) and FRAIL Scale in predicting all-cause mortality within 1 year of discharge in older inpatients at Peking Union Medical College Hospital.

【Methods】 Patients (≥70 years with comprehensive geriatric assessment) discharged from the Department of Geriatrics of Peking Union Medical College Hospital from January 2016 to December 2021 were included. Patients with length of stay ≤24h or >60d were excluded. All-cause mortality within 1 year of discharge were collected from medical files and telephone interviews between August 2022 and February 2023. For statistical analysis, the area under the curve (AUC) was compared by the Delong test, and calibration curves as well as decision curves were plotted.

【Results】 832 patients were included, with a median age of 77 (74-82) years and 54.3% women. 12% patients died within one year. As modeled scores increased, the 1-year mortality risk increased. The AUC of 4 models were as follows: Walter Index 0.89, CCI 0.80, GPI 0.75, and FRAIL Scale 0.75, indicating that all 4 models discriminating well, with the Walter Index surpassing other models (Delong test P <0.05). The calibration curves showed that except for CCI, the other models were well calibrated. Decision curves showed that all 4 models were clinically useful, with the Walter Index best.

【Conclusion】The 4 prediction models are all useful for  predicting all-cause mortality within 1 year of discharge in older inpatients at Peking Union Medical College Hospital. The integrated disease function model outperformed the comorbidity model, while the Walter Index performed best.

PartⅡ  Development and validation of all-cause mortality prediction models for older patients in the geriatric ward within 1 year of hospital discharge

【Background】The Walter Index was validated and recommended to be used in the prediction model of all-cause mortality within 1 year of hospital discharge for older inpatients of geriatric ward in our study, but frailty and intrinsic ability were not included in the model.

【Objective】To develop a prediction model for all-cause mortality in older inpatients of geriatric ward within 1 year after discharge in China.

【Methods】Inpatients (≥70 years old, with comprehensive geriatric assessment) discharged from the Department of Geriatrics, Peking Union Medical College Hospital (PUMCH) between January 2016 and December 2021 were consecutively included. Patients with hospital stays ≤24 hours or >60 days were excluded. Participants were split into training and internal test cohorts (7:3 ratio). Inpatients from three other tertiary hospitals' Department of Geriatrics between January 2019 and December 2021 constituted an external test cohort. Data collection was conducted followed by telephone follow-ups. Least absolute shrinkage and selection operator method and Logistic regression were used to develop prediction models. To calculate the area under the curve (AUC), draw calibration curves and decision curves to assess the prediction efficiency of the model.

【Results】 Training cohort: 582 patients (median age: 78 years; 44.5% female); Internal test cohort: 250 participants (median age: 77 years; 52.8% female); External test cohort: 231 patients (median age: 82 years; 47.2% female). The 1-year mortality for the three cohorts were 12.5%, 10.8%, and 6.1%, respectively. Two models were developed, both including FRAIL, hemoglobin, and albumin. Model 1 included the Charlson Comorbidity Index (PUMCH- Geriatric prognostic index, PUMCH-GPI), and model 2 included metastatic tumors (PUMCH-GPI-tumor, PUMCH-GPIt). The 2 models showed good discrimination, and in the training set PUMCH-GPI AUC = 0.88, sensitivity 83.3%, specificity 80.5 %; PUMCH-GPIt AUC=0.86, sensitivity 76.4%, specificity 83.9%. The calibration curves showed that the calibration curves of PUMCH-GPIt fit the ideal calibration curves in all 3 datasets, and the predicted mortality of PUMCH-GPI were comparable to the actual mortality in the training and the internal validation set, but the calibration curves in the external validation set deviated from the ideal calibration curves.

【Conclusion】We developed 2 all-cause mortality prediction models (PUMCH-GPI and PUMCH-GPIt) for older patients in geriatric ward in China within 1 year of discharge from hospitals, with easy to acquire variables and good discrimination, which are helpful for clinical decision making. More validation of the clinical generalization value is needed.

Part Ⅲ  To compare the efficacy of new models and foreign mortality prediction models for older patients in the geriatric ward within 1 year of hospital discharge

【Background】 To date, there is no prediction model for all-cause mortality in older patients in the department of Geriatrics in China. We developed 2 new models based on Chinese geriatric patients: Peking Union Medical College Hospital Geriatric Prognostic Index (PUMCH-GPI) and PUMCH-GPIt (PUMCH-GPI-tumor) based on patients with metastatic tumors in geriatrics, and their prediction efficacy needs to be compared with international prediction models. 

【Objective】To compare the predictive efficacy of our proposed prediction models, PUMCH-GPI and PUMCH-GPIt, with the Walter Index, Geriatric Prognostic Index (GPI), Charlson Comorbidity Index (CCI), and FRAIL Scale for all-cause mortality within 1 year of discharge from older inpatient in 2 geriatric wards in China, respectively.

【Methods】Patients (Cohort A) consecutively hospitalized in the department of geriatrics of PUMCH from January 2016 to December 2021 (≥70 years old with comprehensive geriatric assessment) were selected, and those with a length of stay ≤24h or >60d were excluded. Patients consecutively hospitalized in the geriatric wards of 3 tertiary hospitals from January 2019 to January 2021 (cohort B) were selected. Physicians were uniformly trained to document and follow up by telephone the prognosis of survival within 1 year of discharge. The prediction efficiency of the model was comprehensively assessed by Delong's test comparing differentiation (AUC), calibration curve comparing calibration, and decision curve analysis evaluating clinical utility.

【Results】 Cohort A included 832 people with a median age of 78 years and 44.5% women; Cohort B included 231 people with a median age of 82 years and 47.2% women. The 1-year mortality rates were 12.0% and 6.1%, respectively. In terms of discrimination, PUMCH-GPI AUC 0.88 (Cohort A), 0.86 (Cohort B); PUMCH-GPIt AUC 0.86 (Cohort A), 0.83 (Cohort B); no less than Walter Index AUC 0.89 (Cohort A), 0.91 (Cohort B), and in Cohort A AUC were significantly better than GPI (0.75), CCI (0.80), and FRAIL scale (0.75). The calibration curves of PUMCH-GPIt were close to the ideal calibration curves in both cohorts, with the best calibration, clinical utility, and overall efficacy better than the other models. The calibration of PUMCH-GPIt in cohort A was better than that of Walter Index as well as the other 3 models, with clinical utility, but the calibration of PUMCH-GPIt was poor in cohort B.

【Conclusion】 Among prediction models for death within 1 year for patients discharged from geriatric units, PUMCH-GPI and PUMCH-GPIt perform well in predicting mortality, no less than Walter Index, but more simple. It is superior to GPI, CCI and FRAIL Scale and may be potentially useful for applications. The number of persons in the external validation set is very limited, and it needs to be validated in more geriatrics departments in more hospitals.