Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification

With population aging and the growing burden of chronic disease, multimorbidity has become an important challenge for clinical decision-making, healthcare resource allocation, and long-term prognosis. Conventional indices such as the Charlson Comorbidity Index, Elixhauser score, FCI, HFRS, and MWI are widely used, but they typically rely on fixed disease lists and weights derived from specific populations. As a result, their calibration and predictive performance may deteriorate across countries, healthcare systems, and age structures, limiting transportable risk stratification. Meanwhile, large language models can integrate heterogeneous health information—including medical history, laboratory measurements, lifestyle, functional status, and social context—providing a computational approach to approximate clinician-like cross-system reasoning. Compared with simple disease counting, LLMs may transform multidimensional health signals into more holistic phenotype representations without relying on fixed weighting schemes. “However, medical LLMs also face concerns regarding reliability, bias, and cross-population generalizability, making large-scale external validation essential.” said the author Jinghao Liang, a researcher at Guangzhou Medical University, “Against this background, we aimed to develop an LLM-based multimorbidity scoring framework for more transportable and interpretable mortality risk stratification and human-machine clinical assessment.”

This study developed and validated an LLM-based multimorbidity scoring system, LLM-MMS, for all-cause mortality risk stratification. The analysis integrated 9 international longitudinal cohorts—UK Biobank, NHANES, SHARE, HRS, CHARLS, CLHLS, ELSA, KLoSA, and MHAS—comprising 770,439 participants. Demographic characteristics, lifestyle factors, examinations, biomarkers, and medical histories were first harmonized across cohorts using a unified data dictionary and converted into standardized English health narratives. DeepSeek-V3 was then used as the core reasoning engine with a fixed prompt template and chain-of-thought prompting, without model fine-tuning, to generate 17 health phenotypes, including overall biological age, multiple organ-specific biological ages, frailty age, and multidimensional grades covering psychological health, diet, behaviors, medical history, income, and family history. These phenotypes were entered into Cox models for all-cause mortality prediction. The model was developed in UK Biobank and externally validated in the other 8 cohorts, with performance compared against conventional indices including CCI, Elixhauser, FCI, HFRS, and MWI using the C-index, calibration, Brier score, and decision-curve analysis. Portability and reproducibility were further tested using alternative LLM backbones and repeated inference runs, while proteomic profiling, model-generated reasoning summaries, and input-masking analyses were used to examine biological plausibility and clinical interpretability.

The results showed that LLM-MMS achieved strong transportability for all-cause mortality risk stratification in people with multimorbidity. In UK Biobank, the model reached a C-index of 0.81, substantially higher than the best conventional comorbidity index at 0.61. Across the other 8 external cohorts, C-indices ranged from 0.67 to 0.86, exceeding recalibrated conventional indices by 0.12 to 0.21, while observed-to-expected ratios generally remained between 0.94 and 1.04, indicating stable calibration. Stratified analyses further showed relatively consistent discrimination across sex and age groups, with particularly clear advantages in adults younger than 60 years, where traditional scores often exhibited floor effects. LLM-MMS also outperformed survival models trained directly on the same raw structured variables in UKB, suggesting that its gain was not simply due to richer input information. Performance varied only modestly across different base LLMs and remained highly reproducible across repeated runs. Proteomic analyses showed that high-risk individuals had molecular signatures related to inflammation, tissue injury, and metabolic stress, while input-masking experiments identified smoking, inflammatory markers, systolic blood pressure, BMI, and HbA1c as major contributors to the final risk score, supporting the clinical and biological plausibility of the model.

The significance of this study lies in introducing a multimorbidity risk-assessment strategy that moves beyond conventional “disease counting plus fixed weights” by using a large language model to integrate cross-organ and multidimensional health information into an interpretable LLM-MMS phenotype. Compared with traditional comorbidity indices, LLM-MMS maintained more stable discrimination and calibration across countries, age groups, and sexes without cohort-specific retraining, suggesting that LLM-derived integrated health representations may offer improved transportability across heterogeneous populations. Proteomic and input-masking analyses further supported its biological and clinical plausibility, linking high-risk phenotypes to inflammation, tissue injury, metabolic stress, and established risk factors such as smoking, blood pressure, BMI, and HbA1c. However, LLM-MMS functions more like a broad prognostic health phenotype than a purely diagnosis-based multimorbidity score, and the inclusion of socioeconomic and other non-diagnostic domains raises questions about fairness and appropriate use. The framework also remains influenced by its UK Biobank development anchor, English-language textification, and potential LLM-related bias. “In the future, we will conduct prospective clinical validation, non English environment testing, continuous calibration and fairness audits, and establish model version management and post deployment monitoring mechanisms to promote its safe use for human-machine collaborative risk stratification.” said Jinghao Liang.

Authors of the paper include Jinghao Liang, Ying Liu, Wei Wang, Zixian Xie, Yijian Lin, Jihao Qi, Yiwen Cai, Yongxuan He, Yaowen Liang, Yiling Chen, Suzheng Zhao, Ziqiu Cheng, Fayuan Wu, Haonan Zhao, Xinhao Liang, Kaiqi Zhang, Yuanlin Yuan, Jingchun Ni, Xiaoyao Lei, Liangyi Yao, Yuanqin Liu, Zishan Huang, Dianhan Lin, Weiqiang Yin, Hengrui Liang, Zhihua Guo, Qiu Wei, Feiying He, Mingyang Jiang, Nanshan Zhong, Qiong Liang, and Jianxing He.

This work was supported by the Natural Science Foundation of Guangdong Province (Grant No. 2026A1515011775), the National Natural Science Foundation of China (Grant No. 82503973), and the R&D Program of Guangzhou National Laboratory (Grant Nos. SRPG22-017 and 2023ZD0519700).

The paper, “Bio-Inspired Artificial Intelligence Multimorbidity Score for Human–Machine Clinical Risk Stratification” was published in the journal Cyborg and Bionic Systems on Sep 30 2026, at https://doi.org/10.34133/cbsystems.0691.

---

---