# Artificial Intelligence Hybrid Survival Assessment System for Robot-Assisted Proctectomy: A Retrospective Cohort Study

Researcher: Ge Zhang
Chinese name: 张格
ORCID identity anchor: https://orcid.org/0000-0002-3116-3246
Canonical researcher: https://drgezhang.com/#person

Journal: JCO Precision Oncology
Year: 2024
Type: Article
DOI: 10.1200/po.24.00089
Canonical page: https://drgezhang.com/papers/doi-10-1200-po-24-00089.html
Markdown record: https://drgezhang.com/papers/doi-10-1200-po-24-00089.md

## Full Authors

1. Shiqian Zhang
2. Ge Zhang
3. Ming Wang
4. Song-Bin Guo
5. Fuqi Wang
6. Yun Li
7. Kaisaierjiang Kadier
8. Zhaokai Zhou
9. Pengpeng Zhang
10. Hao Chi
11. Chuchu Zhang
12. Quanbo Zhou
13. Pin Lyu
14. Shuaiya Zhao
15. Shuaixi Yang
16. Weitang Yuan

<a id="paper-discovery"></a>
## Research & review context

### Research fields

- Artificial intelligence in oncology
- Rectal cancer prognosis

### Review topics

- Machine learning for postoperative survival
- Interpretable cancer risk stratification

### Scientific role

Interpretable survival-modeling framework for rectal-cancer patients after robot-assisted proctectomy.

### Evidence position

Two-center retrospective human prognostic modeling with external evaluation; not prospective clinical utility or a causal survival advantage of robotic surgery.

### Related concepts

- Postoperative prognostic modeling
- Externally evaluated survival models

### Possible citation contexts

- Methods: interpretable AI for postoperative oncology prognosis.
- Evidence synthesis: survival models in the robot-assisted rectal-cancer surgery population, retaining external-performance limitations.

Possible literature-review contexts, not evidence of existing citations. Read the findings and Evidence Scope before citing the original article.

## Evidence Snapshot

**Artificial Intelligence Hybrid Survival Assessment System for Robot-Assisted Proctectomy: A Retrospective Cohort Study**

RAP-AIscore is an interpretable postoperative overall-survival model developed in a retrospective two-center study of 506 patients undergoing robot-assisted proctectomy, with internal validation and independent retrospective SYEMH evaluation.

| Evidence | Value |
| --- | --- |
| Unique total | n=506 |
| NSCVRCS Discovery | n=268; subset of NSCVRCS Meta |
| NSCVRCS Replication | n=116; subset of NSCVRCS Meta |
| NSCVRCS Meta | n=384 |
| SYEMH | n=122 |
| Initial predictors | 53 |
| Algorithms | 12 |
| Modeling schemes | 132 |
| Final predictors | 10 |
| average C-index | 0.833 |

## Research Question

Can a systematically selected hybrid survival-machine-learning framework provide interpretable and externally validated postoperative overall-survival risk stratification for patients with rectal cancer undergoing robot-assisted proctectomy?

## Author Evidence Summary

This retrospective two-center study evaluated 506 unique patients undergoing robot-assisted proctectomy for rectal cancer. The NSCVRCS Meta cohort included 384 patients divided into development and internal-validation subsets; an independent SYEMH cohort included 122 patients. A comparison of 132 candidate survival-modeling schemes selected Surv.coxboost + Surv.aorsf and yielded an interpretable model with 10 predictors. RAP-AIscore showed internal time-dependent survival discrimination and separated patients into three prognostic states. Independent SYEMH evaluation was retrospective: its 1-year AUC was 0.587, with 2- and 3-year AUCs of 0.754 and 0.810. Prospective clinical utility and improved survival from model-guided care remain unestablished.

## Key Findings

<a id="kf1"></a>
### KF1: The study constructed RAP-AIscore using a systematic comparison of 132 candidate survival-modeling schemes in a double-center retrospective cohort.

Context: 506 unique RAP patients; NSCVRCS Meta n=384 plus independent SYEMH n=122. Discovery and Replication are subsets of Meta and must not be counted twice.

| Evidence | Value |
| --- | --- |
| total unique patients | 506 |
| NSCVRCS Meta | 384 |
| Discovery | 268 |
| Replication | 116 |
| SYEMH | 122 |
| modeling schemes | 132 |

Source locator: Abstract: Methods; Methods: Data Sources and Study Population; Results: Outcome and Treatment of Missing Values

<a id="kf2"></a>
### KF2: The selected Surv.coxboost + Surv.aorsf hybrid scheme achieved the highest average C-index among the 132 evaluated schemes.

Context: Systematic retrospective comparison of candidate survival-modeling schemes.

| Evidence | Value |
| --- | --- |
| average C-index | 0.833 |

Source locator: Results: Comprehensive Development of AI-Hybrid Score System; Fig. 2B-D

<a id="kf3"></a>
### KF3: RAP-AIscore showed time-dependent postoperative overall-survival discrimination in the internal NSCVRCS cohorts.

Context: Retrospective internal development and validation; Meta contains the Discovery and Replication subsets.

| Evidence | Value |
| --- | --- |
| Replication 12/36/60/72-month AUCs | 0.851 / 0.849 / 0.902 / 0.915 |
| Meta 12/36/60/72-month AUCs | 0.911 / 0.924 / 0.931 / 0.948 |
| Discovery 12/36/60/72-month AUCs | 0.927 / 0.960 / 0.949 / 0.962 |

Source locator: Results: The Predictive Value of RAP-AIscore System; Fig. 2E-H; Data Supplement Fig. S6

<a id="kf4"></a>
### KF4: The final interpretable RAP-AIscore used 10 accessible predictors.

Context: Predictive model variables; their inclusion does not establish causal effects on mortality.

| Evidence | Value |
| --- | --- |
| predictor count | 10 |
| predictors | pTNM, ROM, CEA, Ns, PostCEA, Ts, CA199, HLP, Postchem, Minutes |

Source locator: Fig. 3C-D; Data Supplement Table S5

<a id="kf5"></a>
### KF5: Conditional-inference analysis separated patients into three prognostic risk states.

Context: The Results text reports Replication low/intermediate/high counts of 68/20/19, which sum to 107 and do not reconcile with the stated Replication cohort size of 116. These subgroup counts are retained as article-reported values; the denominator discrepancy is unresolved and must not be interpreted as documented exclusions or a complete partition of all 116 patients.

| Evidence | Value |
| --- | --- |
| low | RAP-AIscore <= 2.394 |
| intermediate | > 2.394 to <= 5.55 |
| high | > 5.55 |
| Article-reported Replication low/intermediate/high counts (sum 107; cohort 116; unresolved denominator discrepancy) | 68 / 20 / 19 |
| Meta low/intermediate/high counts | 258 / 60 / 66 |
| three-state discrimination | P < 0.001 in Replication and Meta |

Source locator: Results: Interpretability and Clinical Stratification of RAP-AIscore; Fig. 4A-E

<a id="kf6"></a>
### KF6: RAP-AIscore retained prognostic discrimination in the independent SYEMH external cohort, although early external discrimination was modest.

Context: Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.

| Evidence | Value |
| --- | --- |
| 1-year AUC | 0.587 |
| 2-year AUC | 0.754 |
| 3-year AUC | 0.810 |
| high-vs-low survival | P < 0.0001 |
| three-state discrimination | P < 0.001 |

Source locator: Results: Extension to External Cohort From the RAP-AIscore; Fig. 5A-J

## Study Design & Model Development

### Study profile

- Study Design: Double-center retrospective cohort study with internal and independent external validation
- Evidence Type: Human postoperative prognostic survival-modeling study
- Unique total: 506
- Population: Patients with rectal cancer undergoing robot-assisted proctectomy
- Primary Endpoint: In-hospital and post-discharge overall survival
- Secondary Endpoint: Postoperative prognostic risk stratification and time-dependent survival discrimination
- External validation: Yes
- Data modalities: electronic health records, clinical and laboratory variables, operative and pathologic characteristics, longitudinal survival outcomes

### Cohort hierarchy

| Cohort | Role | n | Relationship |
| --- | --- | ---: | --- |
| NSCVRCS Discovery | model development | 268 | Subset of NSCVRCS Meta |
| NSCVRCS Replication | internal validation | 116 | Subset of NSCVRCS Meta |
| NSCVRCS Meta | complete in-house/modeling cohort | 384 | Contains NSCVRCS Discovery, NSCVRCS Replication |
| SYEMH | independent external validation | 122 | Independent cohort |

### Model development

- Initial predictors: 53
- Algorithms: 12
- Modeling schemes: 132
- Selected algorithms: Surv.coxboost, Surv.aorsf
- Feature selection: Filter + sequential-forward-floating wrapper feature selection maximizing Harrell C-index
- Final predictors: 10
- Validation Strategy: 10 repetitions of 10-fold cross-validation, Monte Carlo cross-validation: 100 iterations, sampling rate 0.7, bootstrap: 1,000 iterations, sampling rate 0.7

### Final predictor set

1. pTNM
2. ROM
3. CEA
4. Ns
5. PostCEA
6. Ts
7. CA199
8. HLP
9. Postchem
10. Minutes

### Interpretability

- SHAP
- nomogram
- conditional inference tree
- t-SNE risk-state visualization

## What This Study Adds

- Retrospective development of an interpretable survival model for patients undergoing robot-assisted proctectomy.
- Systematic comparison of hybrid survival-modeling schemes.
- Internal time-dependent discrimination and independent retrospective external evaluation.
- Three-state postoperative prognostic stratification.

## Evidence Scope

### Supports

- Retrospective development of an interpretable RAP-specific survival model.
- Systematic hybrid model selection.
- Internal time-dependent survival discrimination.
- Independent retrospective external evaluation, with modest 1-year AUC of 0.587.
- Three-state prognostic stratification.

### Does Not Establish

- Prospective clinical utility.
- Improved survival from RAP-AIscore-guided care or treatment efficacy.
- Causal effects of the 10 predictors.
- Universal generalizability to all RAP populations or settings.
- Replacement of clinician judgment or pathologic staging.

### Limitations

- The two-center study and independent external evaluation were retrospective.
- Prospective clinical utility and survival benefit from model-guided management were not established.
- External 1-year discrimination was modest, with AUC 0.587.
- Generalizability to all hospitals and RAP populations was not established.
- The Results text reports Replication low/intermediate/high counts of 68/20/19, which sum to 107 and do not reconcile with the stated Replication cohort size of 116. These subgroup counts are retained as article-reported values; the denominator discrepancy is unresolved and must not be interpreted as documented exclusions or a complete partition of all 116 patients.

<a id="citation-use-cases"></a>
## When This Study Is Useful to Cite

### CU1: Which studies developed and externally evaluated survival models specifically for patients undergoing robot-assisted proctectomy?

The study constructed RAP-AIscore using a systematic comparison of 132 candidate survival-modeling schemes in a double-center retrospective cohort. 506 unique RAP patients; NSCVRCS Meta n=384 plus independent SYEMH n=122. Discovery and Replication are subsets of Meta and must not be counted twice. RAP-AIscore retained prognostic discrimination in the independent SYEMH external cohort, although early external discrimination was modest. Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.

Evidence: [KF1](#kf1), [KF6](#kf6)

### CU2: How can systematic hybrid model selection be applied to postoperative survival prediction?

The study constructed RAP-AIscore using a systematic comparison of 132 candidate survival-modeling schemes in a double-center retrospective cohort. 506 unique RAP patients; NSCVRCS Meta n=384 plus independent SYEMH n=122. Discovery and Replication are subsets of Meta and must not be counted twice. The selected Surv.coxboost + Surv.aorsf hybrid scheme achieved the highest average C-index among the 132 evaluated schemes. Systematic retrospective comparison of candidate survival-modeling schemes.

Evidence: [KF1](#kf1), [KF2](#kf2)

### CU3: What evidence supports time-dependent survival discrimination by RAP-AIscore?

RAP-AIscore showed time-dependent postoperative overall-survival discrimination in the internal NSCVRCS cohorts. Retrospective internal development and validation; Meta contains the Discovery and Replication subsets. RAP-AIscore retained prognostic discrimination in the independent SYEMH external cohort, although early external discrimination was modest. Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.

Evidence: [KF3](#kf3), [KF6](#kf6)

### CU4: Which clinical variables were incorporated into the interpretable 10-predictor RAP-AIscore?

The final interpretable RAP-AIscore used 10 accessible predictors. Predictive model variables; their inclusion does not establish causal effects on mortality.

Evidence: [KF4](#kf4)

### CU5: How did RAP-AIscore define low, intermediate and high postoperative prognostic states?

Conditional-inference analysis separated patients into three prognostic risk states. The Results text reports Replication low/intermediate/high counts of 68/20/19, which sum to 107 and do not reconcile with the stated Replication cohort size of 116. These subgroup counts are retained as article-reported values; the denominator discrepancy is unresolved and must not be interpreted as documented exclusions or a complete partition of all 116 patients.

Evidence: [KF5](#kf5)

### CU6: Was RAP-AIscore independently externally validated?

RAP-AIscore retained prognostic discrimination in the independent SYEMH external cohort, although early external discrimination was modest. Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.

Evidence: [KF6](#kf6)

<a id="citation-boundaries"></a>
## What This Study Should Not Be Cited to Claim

- RAP-AIscore-guided management improves patient survival.
- Prospective clinical utility has been established.
- The 10 predictors are causal determinants of mortality.
- Performance generalizes to every hospital or population.
- RAP-AIscore replaces standard clinical or pathological assessment.

<a id="evidence-matrix"></a>
## Evidence Matrix

<a id="evidence-rap-aiscore-1"></a>
### KF1

- Context: 506 unique RAP patients; NSCVRCS Meta n=384 plus independent SYEMH n=122. Discovery and Replication are subsets of Meta and must not be counted twice.
- Finding: The study constructed RAP-AIscore using a systematic comparison of 132 candidate survival-modeling schemes in a double-center retrospective cohort. total unique patients: 506; NSCVRCS Meta: 384; Discovery: 268; Replication: 116; SYEMH: 122; modeling schemes: 132
- Evidence Level: Retrospective survival modeling
- Scope: 506 unique RAP patients; NSCVRCS Meta n=384 plus independent SYEMH n=122. Discovery and Replication are subsets of Meta and must not be counted twice.
- Source Locator: Abstract: Methods; Methods: Data Sources and Study Population; Results: Outcome and Treatment of Missing Values
- Evidence refs: [KF1](#kf1)

<a id="evidence-rap-aiscore-2"></a>
### KF2

- Context: Systematic retrospective comparison of candidate survival-modeling schemes.
- Finding: The selected Surv.coxboost + Surv.aorsf hybrid scheme achieved the highest average C-index among the 132 evaluated schemes. average C-index: 0.833
- Evidence Level: Retrospective survival modeling
- Scope: Systematic retrospective comparison of candidate survival-modeling schemes.
- Source Locator: Results: Comprehensive Development of AI-Hybrid Score System; Fig. 2B-D
- Evidence refs: [KF2](#kf2)

<a id="evidence-rap-aiscore-3"></a>
### KF3

- Context: Retrospective internal development and validation; Meta contains the Discovery and Replication subsets.
- Finding: RAP-AIscore showed time-dependent postoperative overall-survival discrimination in the internal NSCVRCS cohorts. Replication 12/36/60/72-month AUCs: 0.851 / 0.849 / 0.902 / 0.915; Meta 12/36/60/72-month AUCs: 0.911 / 0.924 / 0.931 / 0.948; Discovery 12/36/60/72-month AUCs: 0.927 / 0.960 / 0.949 / 0.962
- Evidence Level: Retrospective survival modeling
- Scope: Retrospective internal development and validation; Meta contains the Discovery and Replication subsets.
- Source Locator: Results: The Predictive Value of RAP-AIscore System; Fig. 2E-H; Data Supplement Fig. S6
- Evidence refs: [KF3](#kf3)

<a id="evidence-rap-aiscore-4"></a>
### KF4

- Context: Predictive model variables; their inclusion does not establish causal effects on mortality.
- Finding: The final interpretable RAP-AIscore used 10 accessible predictors. predictor count: 10; predictors: pTNM, ROM, CEA, Ns, PostCEA, Ts, CA199, HLP, Postchem, Minutes
- Evidence Level: Retrospective survival modeling
- Scope: Predictive model variables; their inclusion does not establish causal effects on mortality.
- Source Locator: Fig. 3C-D; Data Supplement Table S5
- Evidence refs: [KF4](#kf4)

<a id="evidence-rap-aiscore-5"></a>
### KF5

- Context: The Results text reports Replication low/intermediate/high counts of 68/20/19, which sum to 107 and do not reconcile with the stated Replication cohort size of 116. These subgroup counts are retained as article-reported values; the denominator discrepancy is unresolved and must not be interpreted as documented exclusions or a complete partition of all 116 patients.
- Finding: Conditional-inference analysis separated patients into three prognostic risk states. low: RAP-AIscore <= 2.394; intermediate: > 2.394 to <= 5.55; high: > 5.55; Article-reported Replication low/intermediate/high counts (sum 107; cohort 116; unresolved denominator discrepancy): 68 / 20 / 19; Meta low/intermediate/high counts: 258 / 60 / 66; three-state discrimination: P < 0.001 in Replication and Meta
- Evidence Level: Retrospective survival modeling
- Scope: The Results text reports Replication low/intermediate/high counts of 68/20/19, which sum to 107 and do not reconcile with the stated Replication cohort size of 116. These subgroup counts are retained as article-reported values; the denominator discrepancy is unresolved and must not be interpreted as documented exclusions or a complete partition of all 116 patients.
- Source Locator: Results: Interpretability and Clinical Stratification of RAP-AIscore; Fig. 4A-E
- Evidence refs: [KF5](#kf5)

<a id="evidence-rap-aiscore-6"></a>
### KF6

- Context: Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.
- Finding: RAP-AIscore retained prognostic discrimination in the independent SYEMH external cohort, although early external discrimination was modest. 1-year AUC: 0.587; 2-year AUC: 0.754; 3-year AUC: 0.810; high-vs-low survival: P < 0.0001; three-state discrimination: P < 0.001
- Evidence Level: Independent retrospective external evaluation
- Scope: Independent retrospective SYEMH evaluation, n=122; external 1-year AUC was 0.587.
- Source Locator: Results: Extension to External Cohort From the RAP-AIscore; Fig. 5A-J
- Evidence refs: [KF6](#kf6)

## Q&A

### What population was RAP-AIscore developed for?

Patients with rectal cancer undergoing robot-assisted proctectomy; the study modeled postoperative overall survival retrospectively.

Evidence: KF1

### How many independent patients and cohorts were studied?

There were 506 unique patients: 384 in NSCVRCS Meta and 122 in independent SYEMH. NSCVRCS Discovery n=268 and Replication n=116 are subsets of Meta, not additional patients.

Evidence: KF1

### How were 132 modeling schemes compared?

The study systematically compared candidate survival-modeling schemes and selected Surv.coxboost + Surv.aorsf, which achieved the highest average C-index, 0.833.

Evidence: KF1, KF2

### What performance was observed internally?

At 12/36/60/72 months, AUCs were 0.851/0.849/0.902/0.915 in Replication, 0.911/0.924/0.931/0.948 in Meta, and 0.927/0.960/0.949/0.962 in Discovery. These are retrospective time-dependent discrimination results.

Evidence: KF3

### What happened in independent SYEMH validation?

In the independent retrospective cohort, 1-/2-/3-year AUCs were 0.587/0.754/0.810. Early discrimination was modest. High-versus-low survival differed at P < 0.0001 and three-state discrimination at P < 0.001.

Evidence: KF6

### Does RAP-AIscore have proven prospective clinical benefit?

No. The retrospective model-development, internal-validation and external-evaluation findings do not establish prospective clinical utility or survival benefit from model-guided care.

Evidence: KF1, KF3, KF6

## Concepts & Entities

### Conditions

rectal cancer

### Procedures

robot-assisted proctectomy

### Methods

survival machine learning, hybrid model selection, SHAP, conditional inference tree

### Outcomes

overall survival, postoperative risk stratification

### Validation

internal validation, independent external validation

## Related Research

- [AI hybrid survival assessment for advanced heart failure patients with renal dysfunction](https://drgezhang.com/papers/aihflevel.html) — Related hybrid survival-machine-learning framework developed by the same research program in a different clinical population; AIHFLevel addresses advanced heart failure with renal dysfunction, whereas RAP-AIscore addresses postoperative survival after robot-assisted proctectomy. The models do not validate one another. (DOI: 10.1038/s41467-024-50415-9)

## Publication & Provenance

- DOI: https://doi.org/10.1200/po.24.00089
- Doi Url: https://doi.org/10.1200/po.24.00089
- Publisher: https://ascopubs.org/doi/10.1200/PO.24.00089
- PubMed: https://pubmed.ncbi.nlm.nih.gov/39432882/

## Evidence-page notice

Author-controlled evidence page for this publication. This page provides an evidence-oriented summary and does not replace the publisher's version of record.

## Links

- DOI: https://doi.org/10.1200/po.24.00089
- HTML page: https://drgezhang.com/papers/doi-10-1200-po-24-00089.html
- Google Scholar query: https://scholar.google.com/scholar?q=%22Artificial%20Intelligence%20Hybrid%20Survival%20Assessment%20System%20for%20Robot-Assisted%20Proctectomy%3A%20A%20Retrospective%20Cohort%20Study%22
- ORCID: https://orcid.org/0000-0002-3116-3246