Claude can detect patterns to monitor in your health data and help you optimize your general condition.
Can AI help us optimize our health and diagnose latent problems? OpenAI, Anthropic, Google… Leaders in the sector are convinced of this and are gradually developing direct integrations to analyze our health data. Issue ? Europe, with (because of?) its numerous regulations, is slowing down the implementation of these functionalities. However, it is possible to have your medical results analyzed in detail by AI to obtain insights on possible improvements to be made in your daily life and detect possible underlying health problems. However, you must agree to transmit your data to American servers…
The goal of the maneuver? Analyze Apple Health data with Claude
My goal with this test is to have the health data stored on my phone, an iPhone, analyzed by Claude from Anthropic. Having had an Apple Watch Series 8 since December 2022, my phone and watch have accumulated over approximately 1,180 days of data. Valuable data which includes, thanks to my connected watch, a wide variety of medical metrics: heart rate, heart rate variability, signs of hypertension, respiratory rate, blood oxygen level, sleep duration and precise phases, wrist temperature… Not to mention activity level, number of steps, gait stability and a large number of indicators of all kinds. The iOS ecosystem remains quiet about the interpretation of this data and leaves the user to interpret it freely.
Out of a spirit of curiosity and in the desire to optimize my overall physical health, I decided to have all of this data analyzed by Claude, one of the most advanced AI in health. The goal is to obtain actionable recommendations and an opinion on my overall state of health and its evolution over the past three years. Since Apple Health does not have native integration with Claude, I must export my health data before transmitting it. For reasons of relevance and practicality, I will give my data not to Claude but to Claude Code, who will be able to cut and analyze them more easily with the many tools at his disposal (yes, Claude Code is not reserved for code). It would also have been possible to use Claude Cowork, but the interface is more cumbersome to use.
I then ask Claude to analyze my health data and produce a report in Word format with my overall state of health according to my data, my underlying trends for three years, and finally practical and quickly actionable advice to improve my general condition.
Exporting from Apple Health
Apple offers only one way to export health data. Simply go to your Apple Health profile (in your photo at the top right) and scroll through the options until you find “Export all Health data”. The application will then take a few minutes to compile all of your health data in a zip archive called “export”. Once transferred to a computer, you must decompress the archive into a folder. The main health data is located in the “export.xml” file.
Use Claude Code as a medical assistant
Once the data is exported, I open Claude Code in the terminal, in the unzipped folder. I then put the AI into plan mode so that it plans before acting. The AI will therefore use its capabilities to “think” about the tools and methods to use to analyze my health data and produce its report. I provide a fairly precise prompt which delimits the framework of action and lists some possible methods for analysis (machine learning is quite relevant for this data, around 2.2 million points).
Prompt:
Analyse avec précision mes données Apple Santé. ANALYTIQUE: Utilise toutes les techniques pertinentes à ta disposition: - Machine Learning (clustering, détection d'anomalies, prédiction de tendances) - Analyses statistiques avancées (corrélations multivariées, analyses de séries temporelles) - Méthodes scientifiques validées pour l'analyse de données biométriques - N'hésite pas à appliquer des modèles sophistiqués si les données le justifient OBJECTIFS D'ANALYSE: 1. Etat de santé global et scoring multidimensionnel 2. Détection de patterns cachés et corrélations non évidentes entre métriques 3. Identification d'anomalies, outliers, et signaux d'alerte potentiels 4. Analyse temporelle: évolution, cycles, saisonnalité, points de rupture 5. Comparaison avec normes médicales/standards population similaire 6. Prédiction de tendances futures basées sur l'historique 7. Analyse de causalité potentielle (ex: impact activité → sommeil → VFC) LIVRABLES ATTENDUS: Rapport Word exhaustif structuré en: - Executive summary (insights majeurs en 1 page) - Méthodologie utilisée (techniques appliquées, pourquoi) - Analyse détaillée par dimension santé (cardio, sommeil, activité, etc.) - Corrélations cross-métriques et découvertes significatives - Visualisations (graphiques évolution, heatmaps, distributions) - Anomalies et points d'attention médicale éventuels - Recommandations actionnables priorisées par impact/faisabilité - Annexes techniques si pertinent EXIGENCES: - Maximum de profondeur analytique possible - Rigueur scientifique dans les conclusions - Transparence sur les limites/incertitudes - Priorisation claire des insights par importance clinique - Recommandations concrètes, mesurables, réalistes Explore librement les données, teste différentes hypothèses, et n'hésite pas à creuser les aspects qui semblent révélateurs.
Claude begins by exploring the files present and all the data structured in XML (nearly 1 GB) and generates an action plan. In its plan, the AI proposes to parse the XML before starting the analysis, then to divide the data into categories (cardio, sleep, activity), to apply machine learning models to detect hidden patterns and anomalies, to analyze the correlations between the different metrics, and finally to generate the Word report with visuals. Everything is structured in 6 steps: data analysis, cleaning, statistical analyses, predictive models, creation of graphs and writing of the final report.
Claude Code’s execution plan (extract from the overall plan):
Etape 1: Infrastructure données
- Parser XML avec iterparse (gestion mémoire)
- Nettoyage (doublons, valeurs aberrantes, timezone)
- Cache Parquet pour réutilisation
Etape 2: Feature Engineering
- Agrégations temporelles (hourly, daily)
- Features sommeil (durée, efficacité, architecture)
- Features cardio (resting HR, HRV, zones)
- Features activité (steps, energy, workouts)
Etape 3: Analyses Core
- Statistiques descriptives
- Corrélations + PCA + clustering
- Anomaly detection
- Comparaison normes médicales
Etape 4: Analyses Avancées
- Décomposition temporelle
- Changepoints
- Granger causality + VAR
- Forecasting Prophet/ARIMA
Etape 5: Visualisations
- Génération tous graphiques (PNG haute résolution)
- Charts interactifs pour exploration
Etape 6: Rapport Word
- Template python-docx
- Insertion graphiques
- Tableaux comparatifs
- Narratif et recommandations
A very comprehensive report, actionable advice
After 45 minutes of processing, Claude generates an exhaustive Word report of no less than 36 pages structured into 12 sections. The document begins with a one-page executive summary with an overall health score and a multidimensional chart. The report includes around twenty automatically generated visualizations: evolution curves with moving averages, heatmaps, seasonal decompositions, and even 30-day forecasts via the models. Each section (cardiovascular, sleep, activity) includes comparisons with medical reference standards, cited with their scientific sources.
Positive side: my post-exercise cardiac recovery is excellent (25 bpm after 1 minute, a reliable marker of good cardiovascular health), I train regularly with 3.9 sessions per week, and my circadian rhythm remains stable. The areas for improvement are clear: a resting heart rate of 75 bpm (high for my age and my activity level), not enough restorative sleep, and above all a signal that deserves attention: 12.4% of my oxygen saturation measurements are below 94%, which could indicate a sleep apnea syndrome, even if consumer sensors are known for their imprecision on this metric.
Claude also identifies that the benefits of my training appear with a 5-day delay, or that my body recovers less well on weekends than during the week, when one would expect the opposite. The AI points to my social jet lag: I go to bed later on weekends, which disrupts my rhythm. Concrete insights that no consumer health application offers natively.
The exercise proved conclusive. Claude produced in 45 minutes a report that was more detailed and actionable than anything Apple Health offers natively after four years of collection. However, you must be willing to look at the report to get the most out of it, the 36 pages generated require analysis time, and certain technical sections require a minimum of medical expertise to interpret them correctly. Above all, this test gives an overview of what AI will be capable of doing natively in a few years, when direct integrations with health applications will be deployed.
Of course, this type of analysis does not replace professional medical advice. The tool can, however, direct you to the right signals to monitor and provide concrete advice to optimize your daily physical condition. A hybrid approach that could well become the norm.
