Early Detection of Motor Frailty in Older Adults

[zaouache]

Student: Narimane Zaouache
Academic Supervisors: Pierre-Yves Rohan, Laure Combourieu
Host Structure: École Nationale Supérieure d’Arts et Métiers (ENSAM), Institut de Biomécanique Humaine Georges Charpak, Paris (France)
Period: June – July 2025

---

Context

This project is part of the development of SFDFM (Functional Score for Motor Frailty Detection), a tool designed for physiotherapists to detect motor frailty early in older adults.
Objective: improve the digital tool and analyze the clinical relevance of the score to make it operational for field use.

---

Methodology

• Data: 89 patients > 65 years old, collected in private physiotherapy practices using the paper version of the SFDFM (Le Ménez, 2024), then digitized.
• Tools: Python (pandas, scikit-learn), R, Beamer/LaTeX.
• Key steps:
  1. Redesign of the digital interface (form → Excel database → secure consultation).
  2. Statistical analysis and modeling (regressions, cross-validation).
  3. Identification of key variables to optimize the questionnaire.
  4. Definition of robust classification thresholds.
  5. Comparison SFDFM vs Fried score.

---

Results

• Key variables identified: orthopedic exams, mobility tests (chair rise, ground rise, unipedal balance), family situation, memory.
• Profile-based optimization: questionnaire can be reduced for certain patient subgroups.
• Robust thresholds defined:
  • ≤ 8: non-frail
  • 9–18: pre-frail
  • $>$ 18 : frail
    → Allow classification of patients into three frailty categories, comparable to Fried score groups.
• Comparability: SFDFM stable between home-based and hospitalized patients, but less sensitive than Fried in detecting certain clinical differences.

---

Skills Acquired

• Applied data science in healthcare: predictive modeling, model validation.
• Machine learning: variable selection, threshold definition.
• Web development: data collection and consultation interface.
• Clinical data processing: structuring, centralization, and analysis.
• Scientific methodology: comparison with validated scores (Fried).
• Scientific communication: oral presentation, written report, outreach.

---

Useful Links

• 🌐 Website of the SFDFM score application

---

Labels

M2, Data Science, Healthcare, Screening, Modeling, Web Development

Illustrations

Confusion matrix
This figure shows the performance of the SFDFM classification with thresholds 8 and 18. It highlights good detection of frail patients (Se = 0.80, Sp = 0.96).

Most influential variables per subgroup
This table presents the most influential variables for each Fried subgroup.

Performance comparison by subgroup
Comparison between full (14 variables) and reduced (6 variables) models.

Global variable importance
This graph shows the global influence of each variable on prediction error.