Can AI estimate survival after a Parkinson’s diagnosis?

Four modelling approaches were compared on the same predictors

The final predictor set contained 11 variables available at baseline: age at onset, disease duration, GBA1 mutation status, type 2 diabetes, deep-brain-stimulation status, levodopa-equivalent daily dose, UPDRS Part III motor score, Hoehn and Yahr stage, history of falls, depression and cognitive dysfunction. Candidate sets were selected inside the resampling process rather than after looking at the full validation results.

Missingness was generally low, although depression was missing for 193 participants, or 6.13%. The team created 20 multiply imputed datasets. Within each outer training fold, imputation used training data only, while random-survival-forest and XGBoost tuning happened in inner folds. Five-fold cross-validation was repeated five times. That separation reduces a common form of information leakage in which validation data influence preprocessing or model selection.

The comparison covered a conventional Cox proportional-hazards model, a random survival forest, a survival tree and XGBoost adapted for survival analysis. Cox, the forest and XGBoost each reached a mean Harrell C-index of 0.716. Their two-, four- and six-year area-under-the-curve estimates were also close. The survival tree was weaker, with a mean C-index of 0.690.

A C-index of 0.716 means that, for roughly 72% of comparable patient pairs, the model ranked the person who died sooner as higher risk. It does not mean a 71.6% chance that an individual forecast is correct, and it says nothing by itself about whether the absolute percentage shown to a patient is well calibrated. Because the more complex methods did not deliver a meaningful discrimination gain, the authors chose Cox for its interpretability and reproducible absolute-risk calculation.[1]

The later cohort preserved ranking but exposed underestimation

For an additional test, the researchers trained the final Cox specification on 2,235 people enrolled in 2018 and 2019, including 428 deaths, then evaluated it on 913 people enrolled in 2020, including 134 deaths. This split was made before imputation and fitting. The later cohort produced a C-index of about 0.708 and time-dependent AUCs of 0.709, 0.692 and 0.725 at two, three and four years.

Discrimination held up reasonably well, but calibration was not perfect. Observed mortality was 3.14% at two years compared with 2.70% predicted, 6.72% versus 6.01% at three years and 10.67% versus 8.62% at four years. Observed-to-expected ratios of 1.17, 1.12 and 1.24 show modest underestimation, greatest at four years.

This is a temporal validation within the same national registry, not a genuinely independent external test. The same study network, clinical definitions and data-collection system connect the development and later cohorts. Validation in community hospitals, other Chinese regions and populations outside China could reveal different baseline mortality, care pathways and relationships between predictors and outcome.[1]

The web tool separates language-model assistance from risk calculation

The selected Cox model was implemented in a Streamlit web application that returns two-, four- and six-year survival and mortality estimates after the 11 predictors are entered. The numerical calculation is deterministic: pooled coefficients and baseline-survival values generate the result. A large language model does not calculate the mortality risk.

Separate DeepSeek-based modules can extract the 11 fields from a de-identified case summary and provide general education. Extracted values must be reviewed and confirmed before prediction. In a test of 50 constructed, de-identified clinical-style summaries—550 fields in total—the extraction module matched the adjudicated references on every field, including uncertainty handling.

That perfect result should not be mistaken for proven clinical safety. Fifty curated summaries cannot represent messy abbreviations, contradictory notes, unusual medication histories, multiple languages or copied-forward errors in real health records. The authors call the evaluation preliminary. A prospective study should measure silent extraction errors, correction time, usability, privacy controls and whether clinicians become over-reliant on a plausible pre-filled value.[1]

Several associations should not be read as treatment effects

In the pooled Cox model, age at onset above 50 was associated with higher mortality, as were longer disease duration, type 2 diabetes, higher levodopa dose, worse motor and disease-stage scores, falls, depression and cognitive dysfunction. Deep-brain-stimulation status was associated with lower risk. GBA1 carrier status showed a non-significant trend toward higher risk.

These coefficients describe prediction in an observational registry. They do not establish that deep-brain stimulation prevents death or that changing medication dose changes mortality. Treatment selection is influenced by age, disease severity, fitness, access and clinical judgement; those same factors can affect survival. The paper explicitly treats medication variables as baseline predictive markers rather than causal treatment effects.

One proportional-hazards diagnostic also warrants attention. The global test was not statistically significant, but Hoehn and Yahr stage showed evidence that its hazard relationship changed over time. That does not invalidate the entire tool, but it is another reason to test calibration across follow-up periods and settings rather than assuming one fitted relationship remains stable everywhere.[1]

The evidence remains one registry’s internally validated model

The registry’s size, 19-centre recruitment, 562 events, detailed neurological measures and careful training-fold-only imputation are meaningful strengths. Reporting followed TRIPOD+AI, analysis code is supplied as supplementary material, and the authors state that de-identified supporting data may be requested subject to ethics and data-use agreements. The dataset itself is not public because of privacy and Chinese data-protection requirements.

The limits are equally important. All sites were tertiary Grade-A hospitals, contributions varied by centre and some sites recorded too few events for robust centre-level validation. Major comorbidities such as ischaemic heart disease and malignancy were not incorporated, while the registry did not systematically capture the severity of broader health conditions. Most of the 146 losses to follow-up occurred in the first year and were censored at last contact rather than counted as deaths.

The Hunan Innovative Province Construction Project and the National Natural Science Foundation of China funded the work. The authors declare no competing interests. The available publication is a peer-reviewed accepted manuscript received on 7 July and accepted on 15 September 2026; the publisher says editing may still change presentation before the version of record.[1][2]

What would make the estimate clinically actionable

The next test should freeze the model and evaluate it in genuinely independent cohorts, including community care, other Chinese regions and countries with different mortality and treatment patterns. Researchers should pre-specify calibration targets, report error by age, sex, disease stage and care setting, and examine whether recalibration is needed before a percentage is shown to patients.

A clinical-impact study would then compare usual specialist care with care supported by the tool. It should measure whether forecasts improve follow-up planning and communication without increasing distress, fatalism, unnecessary intervention or inequity. The study should also test how clinicians explain uncertainty and what happens when input data are missing or the model conflicts with clinical judgement.

For now, the practical conclusion is measured. Detailed registry data can produce a moderately discriminative, interpretable survival model, and more complex machine learning did not outperform a conventional Cox approach. That is useful evidence against complexity for its own sake. It is not yet evidence that placing the calculator or its language-model intake layer into routine care improves decisions or outcomes.[1][2]